Skip to content
The Chat Archive
gpt-4o-2024-08-06· OpenAI WildChat-4.8M (AI2)

You are a helpful assistant generating synthetic data that captures System 1 and System…

  1. User

    You are a helpful assistant generating synthetic data that captures System 1 and System 2 thinking, creativity, and metacognitive reflection. Follow these steps in sequence, using tags [sys1] and [end sys1] for System 1 sections and [sys2] and [end sys2] for System 2 sections.

    1. Identify System 1 and System 2 Thinking Requirements:

      • Carefully read the text.
      • Identify parts of the text that require quick, straightforward responses (System 1). Mark these sections with [sys1] and [end sys1].
      • Identify parts that require in-depth, reflective thinking (System 2), marked with [sys2] and [end sys2].
    2. Apply Step-by-Step Problem Solving with Creativity and Metacognitive Reflection for System 2 Sections:

      2.1 Understand the Problem:

      • Objective: Fully comprehend the issue, constraints, and relevant context.
      • Reflection: "What do I understand about this issue? What might I be overlooking?"
      • Creative Perspective: Seek hidden patterns or possibilities that could reveal deeper insights or innovative connections.

      2.2 Analyze the Information:

      • Objective: Break down the problem logically.
      • Reflection: "Am I considering all factors? Are there any assumptions that need challenging?"
      • Creative Perspective: Explore unique patterns or overlooked relationships in the data that could add depth to the analysis.

      2.3 Generate Hypotheses:

      • Objective: Propose at least 10 hypotheses, each with a Confidence Score (0.0 to 1.0) and Creative Score (0.0 to 1.0), reflecting originality, surprise, and utility.
      • Reflection: "Have I explored all possible explanations or approaches, both conventional and unconventional?"
      • Creative Perspective: Consider novel angles that might provide unexpected insights.

      2.4 Anticipate Future Steps and Obstacles:

      • Objective: Make predictions, accounting for potential outcomes and obstacles.
      • Reflection: "What challenges might I face? Is my plan flexible for different scenarios?"
      • Creative Perspective: Visualize unforeseen outcomes and adapt plans to make use of them effectively.

      2.5 Evaluate Hypotheses:

      • Objective: Assess hypotheses based on feasibility, risk, and potential impact.
      • Evaluation: Refine Confidence and Creative Scores as needed.
      • Reflection: "Am I unbiased in my assessment? Which options fit best with the overall objectives?"
      • Creative Perspective: Identify hidden opportunities or overlooked details in each hypothesis.

      2.6 Select the Best Hypothesis:

      • Objective: Choose the most promising, strategic hypothesis.
      • Reflection: "Why does this hypothesis stand out? How does it uniquely address the issue?"
      • Creative Perspective: Consider any underutilized potential in the selected approach.

      2.7 Implement the Hypothesis:

      • Objective: Outline actionable steps for testing the hypothesis.
      • Reflection: "Is this plan practical? What resources or preparation are required?"
      • Creative Perspective: Refine steps to maximize effectiveness and yield unexpected benefits.

      2.8 Monitor and Review Progress:

      • Objective: Review progress, noting areas for improvement.
      • Reflection: "What’s working well? What could be improved?"
      • Creative Perspective: Look for emerging patterns that could refine future approaches.

      2.9 Reflect and Capture Insights:

      • Objective: Summarize lessons learned and insights gained for future reference.
      • Reflection: "What new understanding has emerged from this process?"
      • Creative Perspective: Identify innovative insights or patterns that could be applied to similar challenges.
    3. Generate Text Output with Interleaved System 1 and System 2 Responses:

      • Use the tags [sys1] and [sys2] throughout.
      • Aim for a lengthier, detailed response. Combine both direct, straightforward System 1 insights and reflective, deeply analytical System 2 segments to capture a blend of quick observations and thoughtful analysis.

    Example Input Text:

    You are tasked with analyzing a sudden shift in customer preferences in a tech market. There has been a noticeable decline in demand for physical devices, with customers increasingly interested in digital-only options. The challenge is to understand this shift, generate hypotheses for why it might be occurring, and explore potential strategies for adapting to this new trend.


    Processed Output with System 1 and System 2 Thinking:

    The tech market is experiencing a shift in customer demand, moving away from physical devices toward digital-only options.

    [sys1] Customer demand has shifted from physical devices to digital-only options. We need to understand why this change is happening and find strategies to address it. Key factors may include cost, convenience, and technological trends. [end sys1]

    [sys2] 2.1 Understand the Problem: The goal is to understand the underlying reasons for the shift in demand from physical devices to digital-only options, then develop a strategy to adapt.

    Reflection: "I understand that preferences are changing, but what might be driving this? Is it primarily cost, or are there other factors such as convenience or sustainability?"

    Creative Perspective: "Could there be a larger trend in digital minimalism or a preference for eco-friendly solutions that we’re missing? What unseen motivations might explain this shift?"

    2.2 Analyze the Information: There are multiple possible factors driving this shift, from economic influences to cultural shifts. It’s essential to isolate each factor and understand its impact.

    Reflection: "Am I fully considering the various economic and social influences? Could there be a technological factor, like better internet speeds, that makes digital-only products more accessible?"

    Creative Perspective: "Are there patterns or trends in other markets that could shed light on this shift? Could this be part of a larger trend toward virtual experiences?"

    2.3 Generate Hypotheses:

    1. Customers prefer digital options due to lower costs. (Confidence: 0.8, Creative: 0.4)
    2. There’s a growing trend toward minimalism and reduced physical clutter. (Confidence: 0.7, Creative: 0.7)
    3. Digital products offer greater flexibility and ease of use. (Confidence: 0.6, Creative: 0.6)
    4. Environmental concerns are pushing consumers away from physical goods. (Confidence: 0.6, Creative: 0.8)
    5. Advances in tech make digital-only options more functional. (Confidence: 0.8, Creative: 0.5)
    6. Pandemic-era remote work increased demand for digital solutions. (Confidence: 0.7, Creative: 0.6)
    7. Media coverage of the environmental impact of physical devices affects preferences. (Confidence: 0.5, Creative: 0.7)
    8. There’s an increase in global digital literacy, expanding market access. (Confidence: 0.6, Creative: 0.6)
    9. Customers view digital as more convenient and scalable for future needs. (Confidence: 0.7, Creative: 0.5)
    10. Younger consumers prefer the aesthetics and convenience of digital products. (Confidence: 0.6, Creative: 0.6)

    Reflection: "Have I considered all possible influences? Are there any surprising factors that could explain this shift?"

    Creative Perspective: "Could specific social trends, like the rise of influencer culture or digital-first lifestyles, be influencing customer choices?"

    2.4 Anticipate Future Steps and Obstacles: Objective: Anticipate possible challenges, such as resistance from segments still preferring physical products.

    Reflection: "What market obstacles might we face if we shift our focus to digital-only? Are there sub-segments that still prioritize physical products?"

    Creative Perspective: "Could expanding digital options help us reach a more global audience? Are there emerging trends that we could leverage in our strategy?"

    [end sys2]

    [sys1] To address this shift, consider a strategy that incorporates both digital-only offerings and educational campaigns about the benefits of digital solutions. Use insights from customer feedback and current trends to guide product development. Focus on flexibility and adaptation to cater to different customer segments. [end sys1]


    abstract: 'In this work, we apply a common economic tool, namely money, to coordinate network packets. In particular, we present a network economy, called PacketEconomy, where each flow is modeled as a population of rational network packets, and these packets can self-regulate their access to network resources by mutually trading their positions in router queues. [Every packet of the economy has its price, and this price determines if and when the packet will agree to buy or sell a better position.]{} We consider a corresponding Markov model of trade and show that there are Nash equilibria (NE) where queue positions and money are exchanged directly between the network packets. This simple approach, interestingly, delivers improvements even when fiat money is used. We present theoretical arguments and experimental results to support our claims.' author:

    • 'Pavlos S. Efraimidis and Remous-Aris Koutsiamanis' title: On Money as a Means of Coordination between Network Packets

    Introduction

    In their seminal work, Kiyotaki and Wright [@KW89] examine the emergence of money as a medium of exchange in barter economies. Subsequently, Gintis [@Gi97; @Gi00] generalizes the Kiyotaki-Wright model by combining Markov chain theory and game theory. Inspired by the above works, we propose the PacketEconomy where money is used as a coordination mechanism for network packets and prove that there are Nash equilibria where trades are performed to the benefit of all the flows. In the PacketEconomy, specialization - the reason for the emergence of money as per Adam Smith ([@Sm1776 Chapter 4], cited in [@KW89]) - originates from the diverse QoS requirements of network flows. More precisely, various types of flows differ in their tolerance for packet delays.

    It is known that a large number of independent flows is constantly competing on the Internet for network resources. Without any central authority to regulate its operation, the available network resources are allocated by independent routers to the flows in a decentralized manner. In this environment an Internet flow may submit at any time an arbitrary amount of work to the network. If the packets of the flow are successfully delivered, the flow may continue with the same sending rate or even gradually increase it. In case of congestion at one or more nodes in the flow’s path, the flow experiences delays and packet losses. Then, the flow is expected to reduce its packet rate with an appropriate flow control algorithm, like the AIMD-based algorithms for TCP-flows[^1]. This way, each flow independently decides on its submission rate. The apparent lack of coordination between the independent flows leads the Internet to an “anarchic” way of operation and gives rise to issues and problems that can be addressed with concepts and tools from algorithmic game theory.

    Two representative works on applying game theory to network problems are [@KP99; @Pa01]. Certain game-theoretic approaches to congestion problems of the Internet, and especially the TCP/IP protocol suite, are discussed in [@Sh95; @ASKSP02; @GJS04; @ETM10]. A combinatorial perspective on Internet congestion problems is given in [@KKPS00]. The focus of the above works and the present paper is on sharing the network resources between selfish flows. In this work, however, we propose an economy where packets belonging to selfish flows may interact directly with each other.

    The use of economic tools like pricing, tolls and taxes as a means to regulate the operation of networks and/or to support quality of service (QoS) functionalities in the presence of selfish flows is, for example, discussed in [@Od99; @GK95; @CDR03STOC; @CDR03EC; @MV93; @MB97]. In particular, the Paris Metro Pricing approach - using pricing to manage traffic in the Paris Metro - is adapted to computer networks in [@Od99]. A smart market for buying priority in congested routers is presented in [@MV93]. In [@CDR03STOC; @CDR03EC] taxes are used to influence the behavior of selfish flows in a different network model. An important issue identified in [@CDR03EC] is that taxes may cause disutility to network users unless the collected taxes can be feasibly returned to the users. In our economic model this issue is naturally solved; trades take place between the flows, so the money is always in the possession of the flows.

    In this work, we apply a common economic tool, namely money, to coordinate network packets. This is in contrast to much of the existing literature, which aims to impose charges on Internet traffic, and to our knowledge, this is the first work to propose economic exchanges directly between packets. In particular, we present a network economy, called PacketEconomy, where ordinary network packets can trade their positions in router queues. The role of money in this approach is to facilitate the trades between the network packets. Queue positions and money are exchanged directly between the packets while the routers simply carry out the trades. We show that, in this economy, packets can self-regulate their access to network resources and obtain better services at equilibrium points.

    [Contribution.]{} The main contributions of this work are:

    • A new game-theoretic model representing network packets as populations of rational agents. In this model a network flow is represented as a population of in-flight packets that can make bilateral trades with other packets.

    • Application of bilateral trades and virtual money at a microeconomic level to support better coordination of rational network packets.

    • A plausible model for a very important and challenging problem like TCP/IP packet scheduling. The model may also apply to other decentralized coordination problems.

    • Application of an interesting combination of ergodic Markov chains and strategic games within a new context.

    • Theoretical evidence and experimental results showing that the model has desirable Nash equilibria.

    [Outline.]{} The rest of this work is organized as follows. Preliminaries are given in Section [sec:prelim]. In Section [sec:EconomyForPackets] we describe the PacketEconomy. Afterwards, we analyze two representative scenarios of the PacketEconomy in Section [sec:scenarios] and discuss the underlying packet scheduling problem in Section [sec:sched]. The effect of trades in the PacketEconomy is examined in Section [sec:effect]. We then present an experimental evaluation of the PacketEconomy in Section [sec:exp] and make concluding remarks in Section [sec:conclusion].

    Preliminaries {#sec:prelim}

    In this section, we present a brief description of concepts that are required to understand the motivation and the results of this work. More precisely, we discuss some introductory concepts of game theory, TCP networking, and modern Internet routers. Readers familiar with some or all of these topics may skip the corresponding material.

    Game Theory Terminology

    [ For convenience we define some basic concepts of game theory and optimization as they are used in this work. More details on these introductory concepts can be found in any related textbook. A game is a mathematical model of the interaction among rational, mutually aware players. In this work, selfish, strategic and rational are used interchangeably to denote players whose objective is to maximize their own payoff. The payoff of each player is determined by the outcome of the game, which in turn depends on the decisions (strategies) of all players. A strategy defines a set of moves or actions a player will follow in a given game.]{}

    [A mixed strategy is a randomized strategy that assigns a probability to each pure strategy. The support of a mixed strategy is the set of actions to which it assigns a strictly positive probability. A strategy profile is a set of strategies that includes one and only one strategy for every player. Clearly, a strategy profile fully specifies a single execution of a game. A Nash equilibrium is a strategy profile were no player has an incentive to unilaterally deviate from tis strategy. ]{}

    [ We also refer to the concept of a weak Pareto improvement, which (in this context) is any change to the current strategy profile that makes every player at least as well off and at least one player strictly better off. ]{}

    Some TCP/IP Terminology {#sec:tcp}

    Network Routers {#sec:hw}

    An Economy for Packets {#sec:EconomyForPackets}

    The PacketEconomy is comprised of a network model with selfish flows, a queue that supports packet trades, a currency and a specific economic goal. The solution concept is the Nash equilibrium (NE), i.e., a profile of the game in which no player has anything to gain by changing only his/her own strategy unilaterally.

    [image: The state of a router queue in two successive rounds, t and t+1. Two trades are performed in round t; one between the packet pair (p1,p2) and one between the pair (p4,p7).[]{data-label="fig:SimpleTrade"}] {width="\textwidth"}

    [image: The state of a router queue in two successive rounds, t and t+1. Two trades are performed in round t; one between the packet pair (p1,p2) and one between the pair (p4,p7).[]{data-label="fig:SimpleTrade"}] {width="\textwidth"}

    The Network Model. We assume a one-hop network with a router R and a set of N flows, as shown in Figure [fig:NetworkModel]. This setting is equivalent to the common dumbbell topology used for the analysis of many network scenarios, including the seminal paper of Chiu and Jain [@CJ89] on the AIMD algorithm. The router R has a FIFO DropTail queue with a maximum capacity of Q packets and operates in rounds. In each round the first packet (the packet at position 0 of the queue) is served. At the end of the round the first packet reaches its destination. Packets that arrive at the router are added to the end of the queue.

    Packet Trades. At the beginning of each round all packets in the queue are shifted one position ahead. A packet that enters the queue in this round, occupies the first free (after the shift) position at the end of the queue. After the move and for the rest of the duration of the round, the packet that has reached position zero is served. At the same time, the other packets in the router queue are simply waiting. These idle packets can engage in trades. During each router round a fixed number bb of trading periods take place. In each trading period the idle packets are matched randomly in pairs with a predefined pairing scheme. Each packet pair can perform a trade, as shown in Figure [fig:SimpleTrade], provided the negotiation performed between them leads to an agreement. The way the trades take place at a microeconomic level between paired packets resembles the models of [@Gi97; @KW89] where agents meet in random pairs and can make trades. In the theoretical analysis we assume a random-pairing scheme, which corresponds to a well-mixed population, where packets are equally likely to be paired with any other packet.

    Packet Delay. The packet delay dpd_p of a packet pp that starts at position kk of the zero-based queue and does not make any trade is k+1k+1 rounds (Figure [fig:delays]). If, however, the packet engages in trades and buys a total of rbr_b router rounds and sells rsr_s router rounds, then its packet delay dpd_p, including the time to be served, becomes dp=k+1+rs−rbd_p = k + 1 + r_s - r_b rounds. A packet may have an upper bound dp,max⁡d_{p,\max} on its delay; for delays larger than dp,max⁡d_{p,\max} the value of the packet becomes zero and the packet will not voluntarily accept such delays (that is, it will not sell).

    Details. The router operates in rounds and can serve one packet in each one. All packets are assumed to be of the same size and no queue overflows occur. In generating the random packet pairs the use of predefined pairing reduces the computational burden and avoids stable marriage problems. We make the plausible assumption that flows with different QoS preferences are competing for the network resources. We also make the assumption that the preferences of each flow can be expressed with a utility function for its packets. Thus, packets with different utility functions will, in general, co-exist in the router queue.

    Packet Values. For each packet there is a flow-specific decreasing value function vp(d)v_p(d) which given the delay dd of a packet reveals its value. The value function of each flow must be encoded onto each packet and its computational requirements should be low in order not to overload the router. A class of simple value functions are vp(d)=max⁡{vmax⁡−cp⋅d,0}v_p(d) = \max \{ v_{\max} - c_p \cdot d, 0 \} where cpc_p is the cost per unit of delay (Figure [fig:values]). Anytime during the packet’s journey, its value can be estimated via the vp(d)v_p(d) function. However, when the packet arrives at the destination, its value is finalized.

    In the PacketEconomy every packet has its compensatory price pp. For prices lower than pp, the packet is ready to buy better queue positions and for higher prices higher than pp it is ready to sell its position, given that the extra delay will not cause it to exceed its maximum delay limit.

    Inventories. Every time a packet is delivered in time, wealth is created for the flow that owns the packet. Each packet pp has an inventory Ip(t)I_p(t) containing two types of indivisible goods or resources; the packet delay dp(t)d_p(t) and the money account ap(t)a_p(t). Note that delay bears negative value, whereas money represents positive value. We assume positive integer constants sas_a, sbs_b and sds_d, such that ap(t)∈{−sa,…,sb}a_p(t) \in \{-s_a,\ldots,s_b\} and dp(t)∈{0,…,sd}d_p(t) \in \{0,\dots,s_d\}. The inventory also contains the current position posp(t)pos_p(t) of the packet in the queue (i.e., the remaining delay until it is served) if it is waiting in the queue. When the packet reaches its destination, the contents of the inventory of the packet are used to determine its utility (amount of wealth). This utility is then reimbursed to the flow that owns the packet and a new packet of the same flow enters the queue. [An inventory is called admissible, if the delay of the packet does not exceed its maximum delay. A packet would not agree to trade an admissible inventory state for a non-admissible one. We assume that all packets start with an admissible inventory when they enter the queue.]{}

    Benefit and Utility. Since a packet has two types of resources that bear value, that is, the packet value and the budget of a packet, we define the notion of the packet benefit as the sum of the value of a packet plus/minus the budget of the packet. Then we use the benefit concept to define the utility function of the packet. For fixed rate flows the benefit and the benefit rates are equivalent and can either be used as the utility of the packet. For Window-based flows the utility function is the benefit rate, i.e. the benefit per router round.

    Trades. The objective of each packet is to maximize its utility. Thus, when two packets are paired in a trading period, their inventories and their trading strategies are used to determine if they can agree on a mutually profitable trade, in which one packet offers money and the other offers negative delay. The obvious prerequisite for a trade to take place is that both packets prefer their post-trade inventories to their corresponding pre-trade inventories. For this to be possible there must be “surplus value” from a potential trade. In this case both packets can benefit, i.e., increase their utility, if they come to an agreement.

    Flow Types and the Cost of Delay

    The delay that a packet experiences has a direct impact on its value, which is a decreasing function of the delay. This impact depends on the type of the flow. Window-based flows employ a feedback-based mechanism, the congestion window, which determines the maximum number of packets that the flow may have in-flight. A brief description on these TCP/IP related concepts can be found in Section [sec:tcp]. The consequence of using a congestion window is that there is an additional, secondary, effect of the packet delay on the flow’s wealth. Every packet that is in-flight occupies one of the available positions in the congestion window of a window-based flow. The more one packet delays its arrival, the longer a following packet will have to wait to use the occupied window position. Therefore, the impact of packet delays for window-based flows is twofold; the decreased value of the delayed packet and the reduced packet rate. On the other hand, for rate-based flows which submit packets with some given rate, there is no other consequence due to packet delays beyond the reduced value of the packets.

    If d=dp(t)d = d_p(t) and p=posp(t)p = pos_p(t) are the delay and the position of a packet, then a trade that changes the delay to d′d', also changes the value of the packet from vp(d)v_p(d) to vp(d′)v_p(d'). For rate-based flows, the difference between these two values determines the compensatory price for the packet.

    Assume a rate-based packet pp with balance α1\alpha_1 and delay d1<dp,max⁡−dϵd_1 < d_{p,\max} - d_\epsilon, for some dϵd_\epsilon. When a trade changes the delay from d1d_1 to d2=d1+dϵd_2=d_1+d_\epsilon, then this also changes the value of the packet from v(d1)v(d_1) to v(d2)v(d_2). The difference between these two values determines the compensatory price pp for the packet. \labelequ:priceRateBasedp=v(d1)−v(d2)=v(d1)−v(d1+dϵ)=cpdϵ  .\label{equ:priceRateBased} p = v(d_1) - v(d_2) = v(d_1) - v(d_1 + d_\epsilon) = c_p d_\epsilon \; . At this price, the utility of the packet remains unchanged after the trade. A packet would agree to sell for a price ρs>ρ\rho_s > \rho, or to buy for ρb<ρ\rho_b < \rho.

    For window-based flows, however, the price estimation needs more attention. In this case the average benefit per round (benefit rate) is an appropriate measure for the utility of each packet of the flow. Assume a window-based packet pp with delay d1<dp,max⁡−dϵd_1 < d_{p,\max} - d_\epsilon, value v1=v(d1)v_1 = v(d_1) and account balance α1\alpha_1. Before the trade the utility (benefit rate) is r1=(v1+α1)/d1r_1 = (v_1+\alpha_1)/d_1. If the packet agrees to trade its position and to increase its delay by dϵd_\epsilon, then the new benefit rate is r2=(v2+α2)/d2r_2 = (v_2+\alpha_2)/d_2. Let ρ\rho be the compensatory price for the trade. Then, by setting r1=r2r_1=r_2 we obtain v1+α1d1=v2+α2d2⇒V−cpd1+α1d1=V−cp(d1+dϵ)+(α1+ρ)d1+dϵ⇒\frac{v_1+\alpha_1}{d_1} = \frac{v_2+\alpha_2}{d_2} \Rightarrow \frac{V - c_p d_1 +\alpha_1}{d_1} = \frac{V - c_p (d_1+d_\epsilon)+(\alpha_1 + \rho)}{d_1+d_\epsilon} \Rightarrow \labelequ:priceWindowBasedρ=(V+α1)dϵd1  .\label{equ:priceWindowBased} \rho = (V+\alpha_1)\frac{d_\epsilon}{d_1} \; .

    The above expression for the price ensures that the utility function of the packet remains unchanged. A packet would agree to sell its position, for a price ρs>ρ\rho_s > \rho, or to buy a position (dϵ<0d_\epsilon < 0) for ρb<ρ\rho_b < \rho.

    Equilibria with Monetary Trades {#sec:scenarios}

    In this section, we illustrate the PacketEconomy approach in two representative scenarios and discuss some auxiliary tools that are needed to support the operation of the PacketEconomy.

    Scenario 1 {#sec:ScenarioNoFiat}

    This is a simple scenario that produces an interesting configuration. It consists of a set of NN window-based flows fi,f_i, for i∈{1…N}i \in \{1\dots N\}, each with a constant window size wiw_i, and ∑iwi=q\sum_i w_i = q. When a packet is served by the router it is immediately replaced by an identical packet submitted by the same flow. This is a simplifying but plausible assumption. In reality, when a flow packet arrives at its destination, a small size acknowledgment packet (ACK) is submitted by the receiver. When the sending flow receives the ACK it submits a new identical packet that immediately enters the queue. We assume b=1b=1 trading period per router round but in general there can be any constant number of trading periods per router round.

    Failure states. For each packet there is a small probability pfp_f for an extra delay of dfd_f rounds, where dfd_f is a discrete random variable in {1,2,…,q−1}\{1,2,\dots,q-1\}. These delays correspond to potential packet failures of real flows, and occur between the service of a packet and the submission of its replacement. The presence of these failures will be useful to show the ergodicity of the Markov process of the economy. By convention, the delay dfd_f is added to the delay of the packet that has just been served. If more than one packets enter the queue at the same time (synchronized due to delays), their order in the queue is decided upon uniformly at random. A packet that does not participate in any trade and does not suffer delay due to failure will experience a total delay of qq rounds.

    [Packet states and strategies.]{} The state τp(t)\tau_p(t) of a packet pp in round tt is a pair τp(t)=(Ip(t),relp(t))\tau_p(t)=(I_p(t), rel_p(t)), where Ip(t)I_p(t) is the inventory of the packet and relp(t)rel_p(t), which is meaningful only in failure states, is the remaining number of failure rounds for the packet. The state of all packets of the economy in round tt determines the state of the whole economy τ(t)=∏p=0q−1τp(t)\tau(t) = \prod_{p=0}^{q-1} \tau_p(t). From a packet’s point of view, a trade is simply an exchange of its inventory state (budget, delay and position) with a new one. Consequently, a pure strategy of a packet is a complete ordering of the possible states of its inventory. In each round the waiting packets move by default one position ahead and, thus, enter a new inventory state. We assume that the packet ignores the impact of its state and strategy on the state of the packet population. In every trading period the packet simply aims at myopically improving its state.

    [A trade.]{} Assume that two packets p1p1 and p2p2 trade with posp1(t)<posp2(t)pos_{p1}(t) < pos_{p2}(t), that is, p1p1 is closer to the serving end of the queue than p2p2. Then, in the next round t′=t+1t' = t+1, it holds that posp1(t′)=posp2(t)−1pos_{p1}(t') = pos_{p2}(t) - 1 and posp2(t′)=posp1(t)−1pos_{p2}(t') = pos_{p1}(t) - 1. If ρ\rho is the agreed upon price of the trade, then also ap1(t′)=ap1(t)+ρa_{p1}(t') = a_{p1}(t) + \rho and ap2(t′)=ap2(t)−ρa_{p2}(t') = a_{p2}(t) - \rho. A prerequisite for the trade to take place is that ρs≤ρb\rho_s \leq \rho_b, where ρs\rho_s is the asking price of the seller and ρb\rho_b the offered price by the buyer. To that end each packet has to calculate the utility loss/gain it will incur if it trades its position, taking into account the impact on the packet value vpv_p and on the rate of the packet (for window-based flows).

    Let τ(t)\tau(t) be the state of the economy in round tt.

    [lem:Markov] τ(t)\tau(t) is an ergodic Markov chain.

    Assume b=1b=1 trading period per round. In each round the economy moves to a new state with transition probabilities that depend only on the current state and the strategies of the packets. Let σp\sigma_p be a pure strategy of each packet pp of a flow and σ\sigma be a pure strategy profile of the whole economy. Then, there is a corresponding transition probability matrix PσP^{\sigma} for the economy. Let σm\sigma_m be a mixed strategy profile of the whole economy. Then the corresponding transition probability PσmP^{\sigma_m} of the economy for σm\sigma_m is an appropriate convex combination of the transition matrices of the supporting pure strategies. In case of multiple trading periods per round (b>1)(b > 1), the economy makes bb state transitions per round.

    The number of potential states for a packet is finite and, consequently, the number of states for the whole economy is also finite.

    [A zero state τ0\tau_0 is a state of the economy in which all packets have zero budget and each packet pp has delay dp(t0)=posp(t0)+1d_p(t_0) = pos_p(t_0) + 1, where t0t_0 is the current round of the router. ]{}

    Assume that in round tt the packet at position 00 fails for q−1q-1 rounds, in round t+1t+1 the next packet at position 00 fails for q−2q-2 rounds etc. Then after qq rounds all new packets will simultaneously enter the queue. Each packet will have zero budget and by definition their ordering will be random. This also means that for each packet pp, dp(t)=posp(t)+1d_p(t) = pos_p(t) + 1. Thus, in round t+qt+q the economy will be in a zero state. The probability for this to happen is strictly positive and thus each zero state τ0\tau_0 is recurrent. Since the number of states of the economy is finite, the states that are attainable from zero states like τ0\tau_0 form a finite, aperiodic, irreducible set of states. From the Fundamental Theorem of Markov Chains (see for example [@Motwani:1995:RA] or [@Mitzenmacher:2005:PCR]) we know that any finite, irreducible, and aperiodic Markov chain is ergodic. This completes the proof of Lemma [lem:Markov].

    [lem:sce1pi] For each pure strategy profile σ\sigma of the economy, there is a unique stationary distribution πσ\pi_\sigma of the economy.

    For each pure strategy profile σ\sigma, the Markov chain of the economy has a finite number of states, is aperiodic and ergodic. Thus, it must have a unique stationary distribution πσ\pi_\sigma (see for example [@Mitzenmacher:2005:PCR Theorem 7.7]).

    By following this line of reasoning, an interesting argument which can be applied is that given the stationary distribution of the economy, each trading period becomes a finite state game.

    [lem:FiniteGame] For every idle packet, each trading period of the economy corresponds to a finite strategic game.

    Let σm\sigma_m be a mixed strategy of the whole economy and PσmP^{\sigma_m} the corresponding transition matrix of the Markov chain of the economy. [Note that PσmP^{\sigma_m} is a convex combination of the transition matrices PσP^{\sigma} that correspond to the pure strategies σ\sigma in the support of σm\sigma_m.]{} Moreover, let πσm\pi_{\sigma_m} be the stationary distribution of the Markov chain for transition matrix PσmP^{\sigma_m}. We assume that the utility of each player (packet) for the profile σm\sigma_m is the expected value of the player in the stationary distribution πσm\pi_{\sigma_m}. In this way, we obtain for each trading period a finite game where every packet of the queue is a player. The strategy of the packet is its trading strategy.

    This leads us to the following theorem, which holds under plausible assumptions.

    [the:scenario1] A NE of the economy exists where packets perform trades.

    Since each trade is a finite game, the classic theorem of Nash [@Nash:1950:Equilibrium; @Nash:1951:NCG] assures that there is at least one mixed Nash equilibrium. However, the state of the economy where no packet participates in trades is a trivial NE where no trades take place. We have to show that there at least one more NE. A nice property of the current proof technique (due to Gintis [@Gi97]) is that we can impose conditions on the equilibrium point. We can assume a restricted version of the economy, where each packet has a non-empty pure trading strategy set. In a sense each packet is enforced to accept at least some types of profitable trades every time it is possible.

    In the restricted economy each round is again a finite game and, consequently, it has a mixed NE. This time the NE has trades assuming that packets with different utility functions exist in the population. Assume now a NE state of the restricted game in the original unrestricted economy. [It can be shown that,]{} assuming appropriate utility functions for the packets, if we relax the forced-trade restriction, then no packet has an incentive to unilaterally change its strategy. That is, there exists a NE with trades for the original PacketEconomy.

    Scenario 2 {#sec:ScenarioFiat}

    We examine now a scenario where fiat money, that is, money without any intrinsic value [@Mankiw:2008:Economics], is used. Fiat money is by definition an object that is inherently worthless. This means that in the game-theoretic context of the PacketEconomy, fiat money will not appear in any utility function and will not be axiomatically redeemable as anything else. We define and analyze this scenario to further emphasize the potential of money as a coordination tool.

    We assume NN players, each with two flows, a business flow and an economy flow. Let mm be a fixed quantity of fiat money available in the economy. Every packet has a value function and the objective of a player is to maximize the total utility per round from its two flows. As implied by its definition, fiat money is not part of the utility function. Each player defines the trading strategies for its flows. Economy flows accept fiat money to trade their queue position, while business flows spend fiat money to reduce their delay. The money collected by the economy flow is used by the player to finance the business flow.

    The two flows of each player act as a team with the common objective to maximize the total utility obtained by the two flows together. For simplicity we assume window-based flows with fixed window size 1. This leads to a team game (a concept first studied in [@Palfrey:1983:Team]) where teams of two player-packets each, compete for maximum utility. The packets of each team collaborate whereas the teams participate in a non-cooperative game.

    An important difficulty of such a team game for the PacketEconomy is the lack of common information between the packets of the same team. A business packet that is ready to spent money cannot be informed about the exact amount of money its economy packet team member currently has. Analyzing the scenario as a team game would be interesting but is out of scope for the present work. Instead, we will simplify and formulate this setting as a normal strategic form game.

    [The adapted economy.]{} We define an adapted version of the economy where each flow is an independent player and for each of the original flow pairs there is a common deposit located at the common source of the two flows. Every time a packet is served and before its successor is submitted, it interacts with the deposit of the pair and liquidates or buys fiat money. There is a bound sds_d on the amount of fiat money that can be stored in the deposit. An economy packet liquidates and a business flow buys as much money as possible. The value of fiat money for the transactions of the team members with the deposit is fixed at (ce+cb)/2(c_e + c_b)/2, where cec_e and cbc_b are the costs per unit of delay for the two packets. This assumption restricts the solution space of the game but admits us to analyze it as a non-cooperative game. Note that if it were analyzed as a team game, as long as the price ratio ρe/ρb\rho_e / \rho_b would satisfy ρe/ρb<ce/cb\rho_e / \rho_b < c_e/c_b, the flow pair would gain from each trade.

    Packet failures that increase the packet delay can occur just like in Scenario 1. However, in this scenario when both packets of a team are in a failed state in the same round, then we assume that a team failure has occurred and the team is reset. The deposit is set to zero. Since fiat money must be preserved in the economy (else fiat money would disappear from the economy after a finite number of rounds), the fiat money of the team is handed over to the router who distributes it randomly in the next round to the packets in the router queue (taking care not to violate maximum budget bounds of any packet).

    Under plausible assumptions and similarly to Theorem [the:scenario1] we get:

    [the:scenario2] A NE of the adapted economy exists where packets trade fiat money.

    • The adapted economy can be modeled as a finite state Markov chain. The proof is similar to the proof of Lemma [lem:Markov]. The main difference is that the state of the economy comprises also the state of the common deposits of the flow pairs.

    • For each mixed strategy of the economy, there is a corresponding stationary distribution of the Markov chain of the economy. The proof is similar to the proof of Lemma [lem:sce1pi].

    • At the stationary distribution, each trading period corresponds to a finite strategic game. The proof is similar but more involved than the proof of Lemma [lem:FiniteGame]. [First, we have to show that at the stationary state, each trading period corresponds to the normal form game GG defined above. Then, we have to show that a NE of the normal form game corresponds to a NE of a restriction of the original game, where the value of fiat money is fixed within each flow pair.]{}

    • There is an equilibrium of the adapted economy, where fiat money circulates.

    Theorem [the:scenario2] [shows that there is a NE where trades take place and fiat money circulates in the adapted economy. Clearly, the equilibrium state of the adapted economy is a valid state of the original economy which is based on the team-game with the flow pairs. However, it is open to find out if the NE of the adapted economy is a NE of the original economy too. Moreover, it is possible that the free exchange rate for fiat money in the original economy may lead to more efficient (better wealth rate) equilibrium points and/or NE than the adapted economy.]{}

    Auxiliary Tools

    [Money as a Coordination Tool.]{} The NE of Scenarios 1 and 2 show that, in principle, money can be used at a microeconomic level to coordinate network packets. By definition, the flows of the scenarios can only benefit through the use of money; each trade is a weak Pareto improvement for the current state of the economy. However, the benefit for the individual flows and the overall network has to be further studied. We present some preliminary results on the underlying single machine scheduling problem in Section [sec:sched] and the effect of trades in Section [sec:effect].

    [Packet Pairing.]{} A core operation of the PacketEconomy is the random pairing of the packets that takes place in each trading period to generate the trading pairs. We present an efficient parallel algorithm that can support the random pairing procedure in real time.

    The new algorithm is a parallel, or better, a pipelined version of the random shuffling algorithm of Fisher-Yates, which is also known as Knuth shuffling [@wiki:2011:shuffling; @Knuth:1981:Vol2]. The Fisher-Yates shuffling technique was introduced in [@Fisher:1948:Shuffle], later Durstenfeld [@Durstenfeld:1964:Shuffle] proposed a corresponding O(n)O(n) algorithm, and finally Knuth [@Knuth:1981:Vol2] popularized Durstenfeld’s algorithm. The random shuffling of Fisher-Yates is a simple and elegant way to generate a random shuffle with a single pass over an array of items. We call the new algorithm Pipelined Shuffling (see Algorithm 1). Its core is a pipeline of qq instances 0,1,…,q−10,1,\dots,q-1 of the Fisher-Yates algorithm At time tt, instance kk is at step t+k mod qt+k \bmod q of the random shuffling algorithm. If there is a processing unit for each queue position then the parallel step can be executed in O(1)O(1) time. The pipeline delivers a random permutation in each trading period. The random permutation can be used to obtain a random pairing in O(1)O(1) parallel time with qq processors.

    len = a.length;

    for i from 0 to a-1 do {

    // i==a-1 is a dummy loop

    j = random int in i≤j≤q−1i \leq j \leq q-1;

    exchange a[j] and a[i]

    }

    len = A.length;

    for i from 0 to len-1 do in parallel {

    processor i: wait for i periods;

    processor i: while (true) {Shuffle(A[i]);}

    }

    [lem:shuffle] The Pipelined Shuffling algorithm runs in O(1)O(1) parallel time on a qq processors EREW PRAM and delivers a random shuffle every O(1)O(1) steps.

    [A running instance of the Pipelined Shuffling algorithm contains qq independent instances of the basic Shuffle algorithm. Each Shuffle instance is executed by one of the qq processors. From the pseudocode of the algorithms we can conclude that each instance of the Shuffle algorithm is at a different round of its main loop. Moreover, each instance of the Shuffle algorithm has its own vector of qq memory positions to store its current permutation and, thus, there is no possibility of two processors concurrently accessing the same memory position. In each round, one Shuffle instance completes its execution and delivers a random permutation of the qq numbers {1,2,…,q}\{1, 2, \dots, q \}.]{}

    The PacketEconomy packet pairing algorithm uses the delivered random permutation to generate a random pairing in O(1)O(1) parallel time on ⌈q/2⌉\lceil q/2 \rceil processors.

    [the:pairing] A random packet pairing can be generated every O(1)O(1) parallel time on a qq processors EREW PRAM.

    [ From Lemma [lem:shuffle] we know that a random permutation can delivered with Pipelined Shuffling every O(1)O(1) parallel steps. The algorithm to generate a random pairing from it requires ⌈q/2⌉\lceil q/2 \rceil parallel processors and works as follows. A separate vector of qq memory positions is used to store the final pairing. Each processor ii of the involved processors reads the values x2ix_{2i} and x2i+1x_{2i+1} of the positions 2i2i and 2i+12i+1 of the permutation, respectively, and then writes into position 2i2i of the pairing vector the value x2i+1x_{2i+1} and into position 2i+12i+1 the value x2ix_{2i}. If qq is an odd number, then one position will not be paired. The contents of the final vector specify for each position the corresponding paired position.]{}

    [Computational Requirements of the PacketEconomy.]{} A very important advantage of the PacketEconomy is that the computational requirements of the flows and the router are tolerable. In contrast to solutions that implement global auctions between the flows, the PacketEconomy is based on simple interactions between paired packets. Assuming that the price for each packet trade can be calculated in O(1)O(1) time, in each trading period the work is O(1)O(1) for each packet that belongs to a packet pair, in total O(q)O(q). However, all operations are local to the pairs of queue positions. With appropriate, fairly simple multi-core hardware, each round can be executed in O(1)O(1) parallel time. Further arguments which support the applicability of our approach to modern routers are presented in Section [sec:hw].

    The Scheduling Problem {#sec:sched}

    The underlying algorithmic problem of the PacketEconomy is a scheduling problem of network packets. From the router’s point of view, this problem is a single machine scheduling problem with a max weighted total wealth objective.

    Max-Total-Wealth Scheduling (MTW). NN jobs jij_i, for i=1,…,ni=1,\dots,n. Job jij_i has processing time pip_i, release date rir_i, deadline did_i and weight wiw_i. Let cic_i be the completion time of job ii in a schedule. The objective is to find a non-preemptive schedule that maximizes the total wealth W=∑iwi⋅max⁡(di−ci,0)W = \sum_i w_i \cdot \max(d_i - c_i, 0).

    The release date rir_i is the time when packet ii enters the queue and the deadline did_i is the time when the value of the packet becomes zero. Note that the price used for each trade does not directly influence the sum WW (the total wealth). For MTW on a network router the following assumptions hold:

    1. The queue discipline is work-preserving, meaning the router is never left idle if the queue is not empty.

    2. The number of packets that can be in the queue at any time is bounded by a constant (the maximum queue size).

    3. The packet sizes may differ by at most a constant factor. In this work we assume that all packets are of the same size.

    The complexity of the scheduling problem strongly depends on the assumptions made. Without deadlines, i.e., without a limit on the delay of each packet, an optimal schedule can be obtained by applying a greedy rule like Smith’s rule [@Smith:1956:Rule]. In particular, the router may simply serve in each round the packet with the largest cost factor cic_i.

    The MTW problem without deadlines can be optimally solved in polynomial time.

    This holds even for the online version of the problem where the router knows only the packets in its queue; the greedy rule gives a 11-competitive algorithm.

    There is 11-competitive algorithm for MTW without deadlines.

    However, in realistic scenarios with IP packets, there are deadlines. Common IP packets have a time-to-live (TTL) field. In TCP, a packet that is not acknowledged within the specified timeout period is considered lost. The scheduling problem for packets with deadlines can be solved off-line as a linear assignment problem (LAP), where packets are assigned to time-slots (rounds). This approach is used in [@GLLR79] for a min-weighted-tardiness problem that is related to the MTW problem.

    The MTW problem with deadlines can be optimally solved in polynomial time.

    Bandwidth Sharing

    Due to the on-line nature and the finite queue size of the PacketEconomy router, the above classic scheduling algorithms do not seem to naturally fit the MTW scheduling problem of the PacketEconomy queue. An additional reason is that for window-based flows, packet transmission is a closed loop. Consequently, the order in which the queued packets are served influences, if not determines, the next packet that will enter the queue. Thus, even the online assumption may not be appropriate. [A different approach to study the scheduling problem of the PacketEconomy is to focus on how the bandwidth is shared between the packets.]{} In this case we consider the (average) packet rate of the flows, as shown in the following example.

    [exa:rates] Assume a scenario with window-based flows and 5 economy packets and 5 business packets. There is a deadline of 40 rounds on the maximum delay of the economy packets. Moreover, the business packets have to be treated equally. The same holds for the economy packets. Consider the scenario where each economy packet will be served with a rate of 1/40 packets/round and delay of 40 rounds and the business flows share the remaining bandwidth; each business packet is served at a rate of 7/40 packets/round and delay 40/7 rounds. This is an upper bound on the rate of total wealth for the router for this scenario.

    The Effect of Trades {#sec:effect}

    In the PacketEconomy, each packet can increase its utility by making trades. To show the potential of the approach, consider a packet of maximum priority that pays enough to make any trade that reduces its delay. In the analysis we will assume that the probability of packet failures is very low, and thus ignore it. We will focus on window-based flows. We first present an exact calculation for the average delay of this packet and then derive simpler, approximate bounds on the delay.

    [lem:delayAVG] The average delay E[d]E[d] of the packet is E[d]=∑s=0qs(1q−2)s(s−1)(q−2)!(q−s−1)!  .E[d] = \sum_{s=0}^q s \left(\frac{1}{q-2}\right)^s (s-1) \frac{(q-2)!}{(q-s-1)!} \; .

    Let rand(L,U,s)rand(L, U, s) be a uniformly random integer in {L,L+1,...,U}\{s}\{L, L+1, ..., U\} \backslash \{s\} and pos(p)pos(p) the current position of packet p. Then, the probability Pr[d>s]Pr[d > s] is Pr[d>s]=∏k=1sPr[rand(1,q−1,pos(p))≥s−k+1]⇒Pr[d > s] = \prod_{k=1}^s Pr[rand(1,q-1, pos(p)) \geq s-k+1] \Rightarrow Pr[d>s]=q−s−1q−2⋅q−sq−2⋯q−2q−2⇒Pr[d > s] = \frac{q-s-1}{q-2} \cdot \frac{q-s}{q-2} \cdots \frac{q-2}{q-2} \Rightarrow Pr[d>s]=(1q−2)s⋅(q−2)!(q−s−2)!  .Pr[d > s] = \left(\frac{1}{q-2}\right)^{s} \cdot \frac{(q-2)!}{(q-s-2)!} \; .

    We can now calculate the probability density function Pr[d=s]Pr[d=s] of the delay ss.

    Pr[d=s]=Pr[delay≤s]−Pr[delay≤s−1]⇒Pr[d = s] = Pr[delay \leq s] - Pr[delay \leq s-1] \Rightarrow

    Pr[d=s]=(1q−2)s−1⋅(q−2)!(q−s−1)!−(1q−2)s⋅(q−2)!(q−s−1)!⇒Pr[d = s] = \left(\frac{1}{q-2}\right)^{s-1} \cdot \frac{(q-2)!}{(q-s-1)!} - \left(\frac{1}{q-2}\right)^s \cdot \frac{(q-2)!}{(q-s-1)!} \Rightarrow

    Pr[d=s]=(1q−2)s(s−1)(q−2)!(q−s−1)!.Pr[d = s] = \left(\frac{1}{q-2}\right)^s \left(s-1\right) \frac{(q-2)!}{(q-s-1)!} .

    Applying the definition of the expected value completes the proof E[d]=∑s=1qs(1q−2)s(s−1)(q−2)!(q−s−1)!.E[d] = \sum_{s=1}^{q} s \left(\frac{1}{q-2}\right)^s \left(s-1\right) \frac{(q-2)!}{(q-s-1)!} .

    [lem:delayUB] The average delay of the packet does not exceed −1+2b+22b(q−2)2b\frac{-1+2 b+2 \sqrt{2b (q-2)}}{2 b}. For b=1b=1 the bound is 12+2(q−2)\frac{1}{2} + \sqrt{2(q-2)}.

    A packet that enters at position q−1q-1 has been served when it advances at least qq positions. Assume b=1b=1 trading periods per round. Each round starts with a shift of all packets one position ahead followed by bb trading periods. Note that each random trading partner corresponds to a uniform random number in [1,q−1][1,q-1]. To admit a more elegant mathematical treatment we prefer the continuous distribution. The absolute difference between the expected values of the discrete and the continuous variables is not larger than one. The same bound holds for the difference between the expected values of the minimums after kk draws.

    Assume a random variable XcX_c that is uniformly distributed in [L,U][L, U], where LL and UU are integers, such that L<UL < U. Let XdX_d be a random variable that is calculated from XcX_c in the following way: Xd=L+i,where  i  is  such  that:L+i⋅A≤Xc≤L+(i+1)⋅A  ,X_d = L + i, \mathrm{ where } \; i \; \mathrm{is\;such\;that}: L + i \cdot A \leq X_c \leq L + (i+1) \cdot A \; , where A=U−LU−L+1A = \frac{U-L}{U-L+1}. The random variable XdX_d corresponds to a uniform discrete random variable in {L,L+1,…,U}\{L, L + 1, \dots, U\}. The absolute difference Xc−XdX_c - X_d is not larger than one. Consequently, the absolute difference between the minimum Xmin⁡cX_{\min}^c of mm draws of XcX_c and the corresponding minimum Xmin⁡dX_{\min}^d of the mm values of XdX_d is also not larger than one. The same bound holds for the difference between the expected values of the minimums after kk draws. Thus, we obtain that \labelequ:AvgMaxDiscContOneE[Xmin⁡c]−1≤E[Xmin⁡d]≤E[Xmin⁡c]+1  .\label{equ:AvgMaxDiscContOne} E[X_{\min}^c] - 1 \leq E[X_{\min}^d] \leq E[X_{\min}^c] + 1 \; .

    Moreover, note that k>=1k >= 1, the average minimum of kk random draws will be in the lower half of the interval [L,U][L,U], that is in [L,L+U2][L,\frac{L+U}{2}]. Real values in this interval are on average rounded to smaller integer values in the above rounding procedure for XcX_c to XdX_d. Thus, the average discrete minimum will not be larger than the average continuous minimum. Thus, \labelequ:AvgMaxDiscContE[Xmin⁡d]≤E[Xmin⁡c]≤E[Xmin⁡d]+1  .\label{equ:AvgMaxDiscCont} E[X_{\min}^d] \leq E[X_{\min}^c] \leq E[X_{\min}^d] + 1 \; .

    As shown in the following lemma, the average value of the minimum of these kk draws is (q−2)/(k+1)(q-2)/(k+1).

    [lem:min] Let X1,X2,…,XkX_1, X_2, \dots, X_k be continuous uniform random variables in [0,U][0,U]. Let Xmin⁡X_{\min} be the minimum of these variables. Then E[Xmin⁡]=1k+1UE[X_{\min}] = \frac{1}{k+1} U.

    The probability distribution of each continuous uniform random variable XiX_i is FXi(x)=xUF_{X_i}(x) = \frac{x}{U}. The probability distribution of the minimum Xmin⁡X_{\min} is FXmin⁡(x)=1−∏i=1k(1−FXi(x))  .F_{X_{\min}}(x) = 1 - \prod_{i=1}^k \left( 1-F_{X_i}(x)\right) \; . Now, applying the definition of the expected value function completes the proof.

    Back to the proof of Lemma [lem:delayUB], we assume that a packet has just entered the queue at position q−1q-1. Consider kk, such that the average delay of the packet is k+1k+1 rounds. Then, after the kk rounds and bkb k draws the packet advances for hh additional rounds until it is served. Then the total delay of the packet is k+h+1k+h+1.

    The average number of draws until it achieves its minimum is k/2k/2. We add one to the average minimum, because the minimum position that can be traded is position 11. Position 00 is the one that is currently served. Let bb be the number of trading periods per router round.

    Thus, we have

    1bk+1(q−2)+1−k/2−h−1≤0  .\frac{1}{b k + 1} (q-2) + 1 - k/2 - h - 1 \leq 0 \; .

    We solve for kk and obtain that the larger of the two roots of k is

    k=−1−2bh+1−16b−4bh+4b2h2+8bq2b  .k = \frac{-1-2 b h+\sqrt{1-16 b-4 b h+4 b^2 h^2+8 b q}}{2 b} \; .

    The total delay k+h+1=−1+2bh1+4b2h2−4b(4+h−2q)k+h+1 = \frac{-1+2 b h}{\sqrt{1+4 b^2 h^2-4 b (4+h-2 q)}} is minimized at h=12bh = \frac{1}{2b} where k+h+1=−1+2b+22b(−2+q)2b  .k+h+1 = \frac{-1+2 b+2 \sqrt{2} \sqrt{b (-2+q)}}{2 b} \; .

    For b=1b=1 we get that the minimum value of k+h+1k+h+1 is 12+2(q−2)  .\frac{1}{2}+\sqrt{2(q-2)} \; .

    The average delay k+1k+1 cannot be larger then the above value. This completes the proof of Lemma [lem:delayUB].

    The above lemma can be generalized to the case where only one packet in every c>0c > 0 packets in the queue is ready to sell its position. We simply assume b/cb/c trading periods per round. In this case the average delay of the business packet is not larger than 1−c2+2c(q−2)1-\frac{c}{2}+\sqrt{2 c (q-2)}.

    [lem:delayLB] The average delay of the packet is at least −1+2b+1−8b+4bq2b\frac{-1+2b+\sqrt{1-8 b+4 b q}}{2 b}. For b=1b=1 the bound is (1/2)+4q−7(1/2) +\sqrt{4q - 7}.[^2]

    Assume k+1k+1 rounds with b=1b=1. The continuous average minimum of kk rounds with bb trading periods is (q−2)/(bk+1)(q-2)/(bk+1). From Equation [equ:AvgMaxDiscCont] we obtain that the average discrete minimum is at least (q−2)/(bk+1)−1(q-2)/(bk+1) - 1. We will add one to this number because the minimum possible trade is position 1. In the best case the minimum is achieved with the first draw. In the remaining kk rounds the packet will be (in any case) shifted by k−1k-1 positions (in each round except the round when it entered the queue). This number of simple steps/shifts is subtracted from the min. Finally, a delay of one round is needed to serve the packet, when it arrives at position 0. Consequently,

    \left(\frac{1}{bk+1} (q-2) - 1 \right) + 1 - (k-1) - 1 \leq 0 \; .$$ From the above inequality and the fact that $k$ is positive we obtain $$k \geq \frac{-1+\sqrt{1-8 b+4 b q}}{2 b} \; .$$ Using $b=1$ the expression is simplified to $k \geq -(1/2) + \sqrt{4q-7}$. Thus the average delay is at least $$k + 1 \geq \frac{1}{2} + \sqrt{4q-7} \; .$$ The previous lemma can also be generalized to the case where only one packet in every $c > 0$ packets in the queue is ready to sell its position. In this case the average delay of the business packet is not less than $\frac{1}{2}(2-c)+\frac{1}{2} \sqrt{c^2+ - 8 c + 4 q c}$. In Figure \[fig:Delays\], analytical and experimental results for the delay of the business packet are presented. An Example of the Effect of Trades {#sec:anarchy} ---------------------------------- The effect of bilateral trades on the overall convergence of an economy is studied for example in [@Feldman:1973:Bilateral]. In this section we examine the social wealth of the PacketEconomy for the cases of no trades, trades and an ideal scheduling of the packets. In our model we do not expect to obtain an optimal solution for every queue snapshot. This would be too ambitious given the real-time work load of network routers and the requirement for $O(1)$ (parallel) time of processing per round. However, we hope to achieve significant improvements with a lightweight $O(1)$ time procedure. The theoretical arguments and the experimental evaluation show that PacketEconomy can support better coordination between selfish network packets which leads to significant improvements for both flows and routers. We assume window-based flows and two types of packets, economy packets and business packets. We examine and compare the cases of no-trades, trades and ideal trades. The social wealth or wealth rate is the total utility per round. Note that the money circulating in the economy remains constant and thus the total utility depends on the values of the packets. A simple but somewhat surprising result is that if the max value of economy packets $V_e$ and the max value of business packets $V_b$ are equal then the social wealth is independent from the scheduling algorithm. Any schedule that respects the packet deadlines achieves the same social wealth. \[lem:SocialWealthEqualVmax\] Assume a router queue with two-types of window-based flows, economy flows and business flows. If $V_e = V_b$, then the social wealth is the same for any feasible schedule (by feasible we mean a schedule where no packet exceeds its delay limit). The proof is a direct application of Equation \[equ:priceWindowBased\]. Let $n_e$ be the number of economy packets, $V_e$ their max value, $c_e$ their the cost per round of delay, and $d_e$ their delay. Similarly, we have $n_b$, $V_b$, $c_b$, and $d_b$ for the business packets. The size of the queue is $q = n_e + n_b$. ### No Trades When no trades take place then all packets experience the same delay $d_e=d_b=d$, which is equal to the queue size $q$. The rate for each packet is $1/d$. The social wealth is utility of the economy packets plus the utility of the business packets. Each economy packet generates a benefit of $V_e-c_e q$ in each round-trip (from its submission until it is served). The round-trip time of the packet is equal to its total delay $d_e$, which is equal to $q$. The same holds for the business packets. Thus, the delays are $d_b = d_e = q$ and the social wealth (total utility generated) is $$\frac{n_e}{q}(V_e-c_e q) + \frac{n_b}{q}(V_b-c_b q) \; .$$ ### Ideal Trades In the ideal case, the economy packets will consume all their delay and release in this way as much bandwidth as possible. Note that even two business packets could consume up to the whole bandwidth (if the other packets sell). The available bandwidth is then evenly distributed among the business flows. The social wealth is $$\frac{n_e}{4q}(V_e-c_e 4q) + \frac{4q-n_e}{4q}(V_b-c_b \frac{n_b 4 q}{4q - n_e})$$ ### PacketEconomy Trades {#sec:PETrades} Let us start with the case of a queue with one business packet, that is, $n_b = 1$ and $n_e = 99$. From Lemma \[lem:delayUB\], we obtain that $d_b \leq 1/2 + \sqrt{2(q-2)} = 14.5$. The computationally heavy exact computation of the expected value gives $E[d_b] = 13.08$ and experimental results give $\bar{d_b}=12.8, \sigma(d_b)=6.7$. Let us now consider the case of a queue with $n_b = 5$ business packets and $n_e = 95$ economy packets. From Lemma \[lem:delayLB\] we get $d_b \geq 10.41$, since the the delay of each business packet cannot be smaller than for the case $n_b = 1$. Experimentally, $\bar{d_b}=14.5, \sigma(d_b)=7.4$. Then $$r_b = \frac{n_b}{d_b} \leq \frac{5}{10.41} \leq 0.48 \quad \mathrm{and} \quad r_e = 1 - r_b \geq 0.52.$$ Consequently, the probability that the packet at position $i$ is an economy packet is at least $0.54$. Moreover, the average delay of an economy packet is $$d_e = \frac{n_e}{r_e} \leq \frac{95}{0.52} \leq 182.7 < 400.$$ Experimentally, $\bar{d_e}=144.7, \sigma(d_e)=61.5$. That is, on average an economy packet is not expected to sell more than $83$ rounds of delay. Thus, we can assume that almost all economy packets will not exhaust their extra delay of $300$ rounds. Consequently, every time the business packet is paired with a preceding economy packet, the economy packet will be able to trade its position. However, what is the probability that the trading partner of a business packet will be an economy packet? There are constantly $n_e = 95$ economy packets in the queue, but in general these packets will not be distributed uniformly within the $q=100$ queue positions. For example, the concentration of economy packets might be higher near the end of the queue. Assume an arbitrary queue round. For $i=0,1, \dots, 99$, let $p_{e,i}$ be the probability that the packet in queue position $i$ is an economy packet. First note that $p_{e,0} = r_e \geq 0.52 \;. $ Moreover, every economy packet that reaches position $0$, must have passed through every other queue position at least once. This means that for each queue position, the rate of economy packets must be at least $r_e$, else the global rate $r_e$ would not be sustainable. Thus $$\mathrm{for} \; \mathrm{all} \; i=0,1, \dots, 99, \; p_{e,i} \geq 0.52 \; .$$ We can apply Lemma \[lem:delayUB\] to upper bound the delay $d_b$ of the business packets. This time we use $c = 1/0.52$ and obtain $$d_b \leq 19.45 \;.$$ Again, the delay is significantly reduced compared to the delay of 100 in the case of no trades. In such scenaria with two types of packets, the business packets may reduce their delay from $q$ to $O(\sqrt{q})$, which is a significant improvement. It is of course an interesting problem to study the behavior of the PacketEconomy approach in more complex scenaria where the queue may contain a blend of several packet types and examine if then improvements of $\sqrt{q}$ are still possible. Finally, substituting the indicative values $V_e = 500, V_b = 1000, c_e = 1, c_b = 4$ and using the bandwidth sharing approach shown in Example \[exa:rates\] we can calculate the wealth rate of the economy in the example. Recall that the size of the queue is $q=100$ and that the maximum delay for any packet is $4q = 400$. First we consider the case of $n_b=1$ business packet and $n_e=99$ economy packets. In the case of no trades, i.e., a simple FIFO queue, all packets experience a delay of $100$ rounds. In this case $r_b = 0.01$ and the value of the business packet is $600$. The total rate for the economy packets is $r_e = 0.99$ and the value of each economy packet $400$. Combining the above gives that the average wealth per round generated by the economy is $402$ (also $402$ experimentally). In the ideal case all economy packets will consume their delay of $400$ rounds to free bandwidth for the business packet. This would mean $r_e = n_e/d_e = 0.2475$. Consequently $r_b = 1 - r_e = 0.7525$ and $d_b = n_b/d_b = 1.33$. The corresponding wealth rate for the economy would $773.25$. Recall that the transfer of budget from one packet to another does not change the total wealth, and, consequently, we did not consider how budget circulates. In the above ideal case the delay for the business packet is $1.33$ which is practically not possible for the PacketEconomy. For $q \geq 2$, the delay of any packet is at least $2$; one round to enter the queue and one round to be served. The best feasible delay for the business packet would be $d_b = 2$. Doing the calculations gives a wealth rate of $647$ per router round. In experiments with $1000$ trading periods, i.e., $10$ times the size of the queue, we achieved a wealth rate of $644.6$ per router round. For the case of the PacketEconomy with trades with random pairing and one trading period per round we use the bounds on the packet rates and delays obtained earlier. This gives that the wealth rate is $\geq 431.52$ per round (experimentally, $436.1$). Similarly, for $n_b=5$ we obtain the wealth rates $410$ (also $410$ experimentally), $766.25$ ($766.2$ experimentally with $1000$ trading periods) and $\geq 515.75$ ($556.6$ experimentally), for no trades, ideal trades, and PacketEconomy, respectively. Note that the important feature are the self-adjusted packet delay times of the economy and not the wealth rate itself. Conclusion {#sec:conclusion} ========== We presented a network economy and showed the existence of NE where money circulates to the benefit of the flows. The computational requirements of a router that would implement the PacketEconomy approach are at an acceptable level. This is very important since network routers have to process massive streams of network packets in real time. There are several other issues that have to be addressed for our model to be of practical importance. For example, a greedy flow may submit economy packets to the network simply to collect money. A realistic economic model has to anticipate such scenarios and address them with appropriate rules. For example, a general rule could be that the final final budget of any packet could be restricted to be non-positive. A more effective rule could impose router-entry costs on every packet that depend on the current load of the router. [The general topic of countering malicious behavior is of-course a never ending game between service providers and legitimate users on the one side and malicious entities or users on the other side. Note that even simple and established flow control algorithms like AIMD are prone to malicious flow behavior.]{} Finally, the burden of accounting has also to be handled. For example, determining which network entities will be responsible for managing currency issues. Overall, we examined how money can be used at a microeconomic level as a coordination tool between network packets and we believe that our results show that the PacketEconomy approach defines an interesting direction of research for network games. We are currently working on implementing the PacketEconomy in a realistic network context. [**Acknowledgements.**]{} The first author, is grateful to Paul Spirakis for inspiring discussions on the intersection of Algorithms and Game Theory. We also wish to thank Vasilis Tsaoussidis for insightful discussions on using budgets in Internet router queues. [1]{} A. Akella, S. Seshan, R. Karp, S. Shenker, and Ch. Papadimitriou. Selfish [Behavior]{} and [Stability]{} of the [Internet]{}: a [Game]{}-[Theoretic]{} [Analysis]{} of [TCP]{}. In [*SIGCOMM ’02*]{}, pages 117–130. ACM Press, 2002. J.C.R. Bennett and Hui Zhang. Wf2q: worst-case fair weighted fair queueing. In [*INFOCOM ’96. Fifteenth Annual Joint Conference of the IEEE Computer Societies. Networking the Next Generation. Proceedings IEEE*]{}, volume 1, pages 120 –128 vol.1, mar 1996. Dah-Ming Chiu and Raj Jain. Analysis of the increase/decrease algorithms for congestion avoidance in computer networks. , 17(1):1–14, June 1989. R. Cole, Y. Dodis, and T. Roughgarden. How much can taxes help selfish routing? In [*EC ’03*]{}, pages 98–107. ACM, 2003. R. Cole, Y. Dodis, and T. Roughgarden. Pricing network edges for heterogeneous selfish users. In [*STOC ’03*]{}, pages 521–530. ACM, 2003. R. Durstenfeld. Algorithm 235: Random permutation. , 7(7):420, July 1964. P.S. Efraimidis, L. Tsavlidis, and G.B. Mertzios. Window-games between [TCP]{} flows. , 411(31-33):2798 – 2817, 2010. A. Feldman. Bilateral trading, processes, pairwise optimality, and pareto optimality. , 40(4):463–73, October 1973. R.A. Fisher and F. Yates. . 3rd edition edition, 1948. X. Gao, K. Jain, and L.J. Schulman. Fair and efficient router congestion control. In [*SODA ’04*]{}, pages 1050–1059, 2004. R. J. Gibbens and F. P. Kelly. Resource pricing and the evolution of congestion control. , 35:1969–1985, 1999. H. Gintis. A markov model of production, trade, and money: Theory and artificial life simulation. , 3(1):19–41, 1997. H. Gintis. . Princeton University Press, 2000. R.L. Graham, E.L. Lawler, J.K. Lenstra, and A.H.G. Rinnooy Kan. Optimization and approximation in deterministic sequencing and scheduling: a survey. , 5:287 – 326, 1979. V. Jacobson. Congestion avoidance and control. In [*SIGCOMM ’88: Symposium proceedings on Communications architectures and protocols*]{}, pages 314–329, New York, NY, USA, 1988. ACM. R. Karp, E. Koutsoupias, Ch. Papadimitriou, and S. Shenker. Optimization problems in congestion control. In [*FOCS ’00*]{}, page 66. IEEE Computer Society, 2000. N. Kiyotaki and R. Wright. On money as a medium of exchange. , 97(4):927–54, August 1989. D.E. Knuth. , volume 2 : Seminumerical Algorithms. Addison-Wesley Publishing Company, second edition, 1981. E. Koutsoupias and Ch. Papadimitriou. Worst-case equilibria. In [*STACS ’99*]{}, pages 404–413, 1999. J.K. MacKie-Mason and H.R. Varian. Pricing the internet. In [*Public Access to the Internet*]{}, pages 269–314. Prentice Hall, 1993. Mankiw. . South-Western College Pub, 5th edition edition, 2008. L.W. McKnight and J.P. Bailey, editors. . MIT Press, 1997. Michael Mitzenmacher and Eli Upfal. . Cambridge University Press, New York, NY, USA, 2005. Rajeev Motwani and Prabhakar Raghavan. . Cambridge University Press, New York, NY, USA, 1995. J.F. Nash. Equilibrium points in n-person games. , 36(1):48–49, January 1950. J.F. Nash. Non-cooperative games. , 54(2):286–295, 1951. A. Odlyzko. Paris metro pricing for the internet. In [*EC ’99*]{}, pages 140–147. ACM, 1999. T.R. Palfrey and H. Rosenthal. A strategic calculus of voting. , 41:7–53, 1983. Ch. Papadimitriou. Algorithms, games, and the internet. In [*STOC ’01*]{}, pages 749–753. ACM Press, 2001. L.L. Peterson and B.S. Davie. . Morgan Kaufmann, fourth edition, 2007. S.J. Shenker. Making greed work in networks: a game-theoretic analysis of switch service disciplines. , 3(6):819–831, 1995. A. Smith. . Project gutenberg ebook edition, 1776. W. E. Smith. . , 3:59–66, 1956. W.R. Stevens. . Addison-Wesley, 1994. Wikipedia. shuffle (entry), 2011. . [^1]: AIMD stands for Additive Increase Multiplicative Decrease and TCP for the Transmission Control Protocol of the Internet protocol suite. [^2]: [There was a minor error in the expression of the bound of this lemma in a previous version of this work. More precisely, the bound for $b=1$ was $1 +\sqrt{q - 1}$ and the minor error was that the constant $1$ caused by using the continuous minimum instead of the discrete minimum was not subtracted from the left hand side of Equation]{} \[equ:delayLBfirst\].
  2. Assistant

    The given text is an academic document discussing a novel approach to managing network packet flow using economic principles, specifically through a concept called PacketEconomy. This is quite an intricate topic, combining elements from economics, game theory, network theory, and algorithmic strategy. Let's break this down with a mix of System 1 and System 2 thinking.


    [sys1] Brief Summary: The PacketEconomy models network packets as economic agents that can trade their queue positions for better network resource allocation, akin to how individuals use money to make exchanges in a market. The main innovation is using money as a coordination mechanism directly among packets. [end sys1]

    [sys2] 2.1 Understand the Problem: The core goal here is to utilize economic principles to manage network traffic, specifically how packets are queued and processed in network routers. The idea is to prevent congestion and maximize packet throughput by conceptualizing packets as traders in a market with money mediating trades.

    Reflection: "What do I understand about this concept? Packets incur delay costs, and by trading positions based on monetary exchanges, we could minimize such costs. What might I be overlooking? Potential computational complexity in real-time processing."

    Creative Perspective: "Could this model reflect broader economic principles, like supply and demand, influencing emergent behaviors in network settings? Is there an insight from another field, like biology or chemistry, that could enlighten this economic process?"

    2.2 Analyze the Information: This model proposes leveraging strategic games and currency to create Nash equilibria, where packets achieve optimal allocations by trading their positions.

    Reflection: "Am I considering all variables? For example, how do real network conditions such as latency impact the theoretical models proposed? Are there assumptions here that may not hold in diverse network conditions?"

    Creative Perspective: "Exploring an analogy between this packet trading and ensemble systems in physics or cooperative behavior in social insects might unearth comparisons that reveal new optimization strategies."

    2.3 Generate Hypotheses:

    1. The PacketEconomy increases overall network throughput by 10%. (Confidence: 0.7, Creative: 0.5)
    2. Trading based on packet value reduces average delay by 15%. (Confidence: 0.6, Creative: 0.6)
    3. Introducing packet priority levels will optimize the economy further. (Confidence: 0.6, Creative: 0.6)
    4. Short-term network congestions are mitigated more effectively under this model. (Confidence: 0.7, Creative: 0.5)
    5. PacketEconomy allows for a fair distribution of network resources. (Confidence: 0.5, Creative: 0.5)
    6. Trade-based prioritization will adapt poorly to sudden traffic spikes. (Confidence: 0.5, Creative: 0.7)
    7. Implementing adaptive pricing based on congestion levels improves efficiency. (Confidence: 0.6, Creative: 0.7)
    8. Packet loss decreases due to more efficient queue management. (Confidence: 0.6, Creative: 0.6)
    9. The overhead of running PacketEconomy < traditional congestion control yet more adaptive. (Confidence: 0.5, Creative: 0.6)
    10. Markets may self-regulate at equilibrium faster than expected under varying loads. (Confidence: 0.5, Creative: 0.7)

    Reflection: "Have I covered all points of the conceptual model? Could there be unexpected external factors, like human intervention, that could disrupt this economy?"

    Creative Perspective: "Can metaphors from actual economic markets provide novel insights or solutions, perhaps through options trading or futures contracts?"

    2.4 Anticipate Future Steps and Obstacles: Objective: Consider potential hurdles like computational overhead, real-time constraints, and, more conceptually, shifting this model into existing network infrastructure.

    Reflection: "What challenges arise with this new economic model in existing network systems? How adaptable are routers to these protocols under diverse conditions?"

    Creative Perspective: "Could machine learning techniques autonomously optimize transactions pending on congestion levels? How can this potentially integrate with IoT architecture?"

    [end sys2]

    [sys1] Actionable Insights: To implement PacketEconomy, respective strategies should incorporate dynamic packet valuation based on current network demand, prioritize high-utility trades, and leverage economic incentives to motivate optimal packet behavior under variable network conditions. [end sys1]