The Rock--Paper--Scissors Game

TL;DR

This paper introduces Rock-Paper-Scissors as a model for decision-making and strategic interactions, explaining key concepts and reviewing recent theoretical and empirical studies on its non-equilibrium dynamics and applications.

Contribution

It provides an accessible introduction to RPS for physics students and reviews recent advances in understanding its complex non-equilibrium behaviors and applications in ecology and economics.

Findings

Collective cycling observed in RPS dynamics

Conditional response patterns influence strategy stability

Microscopic mechanisms promote cooperation in RPS models

Abstract

Rock-Paper-Scissors (RPS), a game of cyclic dominance, is not merely a popular children's game but also a basic model system for studying decision-making in non-cooperative strategic interactions. Aimed at students of physics with no background in game theory, this paper introduces the concepts of Nash equilibrium and evolutionarily stable strategy, and reviews some recent theoretical and empirical efforts on the non-equilibrium properties of the iterated RPS, including collective cycling, conditional response patterns, and microscopic mechanisms that facilitate cooperation. We also introduce several dynamical processes to illustrate the applications of RPS as a simplified model of species competition in ecological systems and price cycling in economic markets.

Equations

$$w_r=w_p=w_s=\frac{1}{3},$$

$$g^0=\frac{1+a}{3}.$$

$$g^{\prime}=g^{0}-\frac{(a-2)n}{2N}\Bigl{[}(w_{r}^{\prime}-1/3)^{2}+(w_{p}^{\prime}-1/3)^{2}+(w_{s}^{\prime}-1/3)^{2}\Bigr{]}\;.$$

$$\frac{1}{3}+\Bigl{[}(w_{r}^{\prime}-1/3)^{2}+(w_{p}^{\prime}-1/3)^{2}+(w_{s}^{\prime}-1/3)^{2}\Bigr{]}\;,$$

$$\frac{\mathrm{d}{\rho }_r}{\mathrm{d}t}$$

$$\frac{\mathrm{d}{\rho }_p}{\mathrm{d}t}$$

$$\frac{\mathrm{d}{\rho }_s}{\mathrm{d}t}$$

$${\rho }_r^{\ast }$$

$${\rho }_p^{\ast }$$

$${\rho }_s^{\ast }$$

$$g_r$$

$$g_p$$

$$g_s$$

$$\frac{\mathrm{d}{\rho }_r}{\mathrm{d}t}$$

$$\frac{\mathrm{d}{\rho }_p}{\mathrm{d}t}$$

$$\frac{\mathrm{d}{\rho }_s}{\mathrm{d}t}$$

$$w_l^0=\frac{mh-3l(h-m)}{ml+(m-l)h},w_m^0=\frac{(m+l)h-3ml}{ml+(m-l)h},w_h^0=\frac{ml-(m-l)h}{ml+(m-l)h}.$$

$$P(s_t^{(y)}|s_{t-1}^{(x)})=\frac{e^{\beta g(s_t^{(y)}|s_{t-1}^{(x)})}}{e^{\beta g(l|s_{t-1}^{(x)})}+e^{\beta g(m|s_{t-1}^{(x)})}+e^{\beta g(h|s_{t-1}^{(x)})}},$$

Full text

Aimed at students of physics with no background in game theory, this paper introduces the concepts of Nash equilibrium and evolutionarily stable strategy, and reviews some recent theoretical and empirical efforts on the non-equilibrium properties of the iterated RPS, including collective cycling, conditional response patterns, and microscopic mechanisms that facilitate cooperation. We also introduce several dynamical processes to illustrate the applications of RPS as a simplified model of species competition in ecological systems and price cycling in economic markets.

Keywords: cyclic dominance; non-cooperative game; decision making; social cycling; conditional response; non-equilibrium

1 Introduction

Statistical mechanics aims at understanding collective behaviours of many-particle systems from microscopic interactions [1]. If the system is a physical one, each interaction among a subset of particles is associated with an energy, and the total energy of the system is simply the sum of all these energies. The system prefers to stay in microscopic configurations that minimize the total energy, but it is constantly disturbed by the environment. This competition between energy minimization and environmental perturbation leads to very rich non-equilibrium dynamics and to many equilibrium phase transitions in the system’s macroscopic property [1].

Various collective behaviours, driven by strategic interactions among selfish agents, also emerge in game systems. As a new research field of statistical mechanics, exploring and understanding the complex non-equilibrium properties of such competitive social systems became rather active in recent years [2, 3, 4, 5, 6, 7, 8, 9]. Yet statistical mechanical approaches to strategic interactions, compared with physical systems, are facing two additional major challenges.

In a game system, each interaction brings a payoff to every involved agent, but a fundamental distinction with physical systems is that the payoffs for different agents of the same interaction are in general different. Such differential payoffs cause all the conflicts and competitions in the system [10]. Every agent seeks to maximize its own payoff, but the increase of one agent’s payoff does not necessarily mean an increase of the sum of payoffs of all the agents. Since the microscopic dynamics is not guided by the total payoff, the conventional concept of equilibrium Boltzmann distribution of the total payoff is not useful.

Another major challenge is that the microscopic mechanisms of decision-making are quite unclear. In many game systems the agents have certain degree of intelligence, and they make decisions in complicated ways taking into account both the past experiences and the anticipated future events. Furthermore, the microscopic parameters of decision-making may evolve in time as a result of learning and adaptation.

Various games have been investigated in the literature, of which the most widely discussed probably is the Prisoner’s Dilemma game, devised by Albert Tucker about 60 years ago [11, 12]. This game is a paradigm for studying cooperation of selfish agents and is the focus of thousands of research papers [13, 14]. On the other hand the Rock-Paper-Scissors (RPS) game is a paradigm for studying competition caused by cyclic dominance [7], yet it is much less discussed. In the present paper we review some aspects of the RPS game for physics students, assuming the reader has no background in game theory.

Being a popular game, the origin of RPS has been difficult to trace, but there is some written evidence suggesting the Chinese played it already in the Han Dynasty more than two thousand years ago. It is the simplest competition system manifesting the ancient cyclic-dominance concept of oriental taoism philosophy. This game brings fun to people of all ages and occasionally serves as a fair mechanism to resolve choice conflicts among friends or family members. There are three possible action choices: $R$ (rock), $P$ (paper) and $S$ (scissors). Action $R$ is better than $S$, which in turn is better than $P$, which in turn is better than $R$ (Figure 1A). Because of this cyclic dominance none of the three actions is an absolute winning choice. In the simplest case the game is played by two players and they compete simultaneously. For example if one player X chooses $R$ while the other player Y chooses $S$, then X is the winner.

The RPS game is not merely a game for fun, it actually has fundamental importance as a basic model system for non-cooperative strategic interactions. In theoretical studies, cyclic dominance is expressed more quantitatively by a payoff matrix. A frequently used one is shown in Figure 1B with the parameter $a$ being the reward of the winning action. For example, if player X chooses action $R$ and her opponent Y chooses $S$ then the payoff to X is $a$ while that to Y is [math]; if both players choose the same action (e.g., $P$ versus $P$) then a tie occurs and each player gets unit payoff. To ensure the property of cyclic dominance we require $a>1$. At the specific value $a=2$ the total payoff of the two players is the same no matter whether the output is win-lose or tie. If $a>2$, win-lose offers a higher total payoff than tie, while the reverse is true for $1<a<2$. In more general payoff matrices the winning payoffs of the three different actions are different, then the rotational symmetry among $R$, $P$, and $S$ are broken (an example will be given in Section 5).

In the following sections, we first introduce the concepts of Nash equilibrium and evolutionarily stable strategy for the RPS game (Section 2), and then discuss some recent theoretical and empirical efforts on the non-equilibrium properties of iterated RPS, such as collective cycling, conditional response patterns, and microscopic mechanisms that facilitate cooperation (Sections 2 and 3). The applications of RPS in understanding species competition of ecological systems and price cycling of economic markets are discussed in Sections 4 and 5. We conclude this review with a brief outlook in Section 6.

2 Nash Equilibrium, Evolutionary Stability, and Cooperation-Trap Strategies

There are two general theoretical frameworks to study strategic interactions, classic game theory [16, 10] and evolutionary game theory [17, 18, 19]. Classic game theory is based on the assumption that the players can make completely rational decisions, while evolutionary game theory tries to understand game outcomes from the angle of evolution and adaptation. Consider a simple scenario of two players X and Y repeating RPS for an indefinite number of rounds. The payoff matrix is identical to that of Figure 1B, with $a$ being a constant. What shall we expect to observe concerning this system’s long-time behaviour? Classic game theory gives a clear-cut answer that it will reach a unique mixed-strategy Nash equilibrium, while evolutionary game theory is more cautious, emphasizing that the actual microscopic dynamics of decision-making is also a crucial factor.

2.1 Nash-equilibrium mixed strategy

Classic game theory assumes that the players all have unbounded rationality, and the solution concept of Nash equilibrium (NE) plays a fundamental role in this theoretical framework. In an iterated RPS between two such players, if player X always chooses the same action (say $R$), the other player Y naturally will always choose the winning action ($P$) to maximize his own payoff. If player X always chooses between two actions (say $R$ and $S$), player Y will also take advantage of such a regularity and beat X by always choosing between actions $R$ and $P$. To avoid being exploited, player X should therefore adopt all the actions with positive probabilities. But how?

Since any regularity of choices might be detected and be exploited by the rational (and intelligent) player Y, it is safe for player X to make an action choice in each game round completely independent of her choices in the previous rounds. Let us denote by $w_{r}$, $w_{p}$ and $w_{s}$ the respective probability of $R$, $P$ and $S$ being chosen by player X in one game round ($w_{r}+w_{p}+w_{s}\equiv 1$). The probability vector $(w_{r},w_{p},w_{s})$ is referred to as a mixed strategy in the game theoretical literature [10]. (If two of the action probabilities are strictly zero, a mixed strategy reduces to a pure strategy.) Under a given mixed strategy of X, if player Y takes action $R$ then his expected payoff $g_{r}$ per game round is simply $g_{r}=w_{r}+aw_{s}$. Similarly the expected payoffs $g_{p}$ and $g_{s}$ of $P$ and $S$ are $g_{p}=w_{p}+aw_{r}$ and $g_{s}=w_{s}+aw_{p}$.

Notice that if player X chooses a particular mixed strategy $(w_{r},w_{p},w_{s})$ with

[TABLE]

the expected payoff of player Y is independent of his action and $g_{r}=g_{p}=g_{s}=g^{0}$ with

[TABLE]

In other words, it is impossible for Y to exploit X if the latter completely randomize her action choices. Similarly if player Y sets his mixed strategy to be $(1/3,1/3,1/3)$, then the expected payoff of player X is equal to $g^{0}$ and it can not be further increased no matter how hard X tries to adjust her mixed strategy. That is, the chance of player X to take advantage of Y is also eliminated.

Equation (1) is a NE mixed strategy for the two-person iterated RPS. In general, for a game with two or more action choices and involving two or more players, we define an action probability vector of a player $i$ as a NE mixed strategy for this player if $i$ is unable to increase its expected payoff by changing to any another mixed strategy when all the other players keep their own mixed strategy. If every player of the population is taking such a mixed strategy the whole system is then said to be in a Nash equilibrium [10].

It can be easily checked that the two-person iterated RPS has only a unique NE at any $a>1$ and, the probability vector $(1/3,1/3,1/3)$ is the only NE mixed strategy. We can extend the discussion to the scenario of more than two players. At each game round every player competes with all the other players or only with a single randomly chosen player. If we consider only probability vectors $(w_{r},w_{p},w_{s})$ that are strictly mixed (satisfying $w_{r}w_{p}w_{s}>0$), then it is relatively easy to prove that $(1/3,1/3,1/3)$ is also the unique NE strictly mixed strategy [20].

The strategy $(1/3,1/3,1/3)$ is maximally random. When a system of $N$ players reaches the Nash equilibrium, all the $3^{N}$ possible microscopic configurations are equally likely to be observed, and the entropy of the system achieves the global maximum value. Since every player makes decisions independent of other players and of the previous decisions, the dynamical property of the system is completely trivial.

The mixed-strategy Nash equilibrium, although being unique for the iterated RPS game, may not necessarily be stable under small perturbations. Since different action choices bring the same expected payoff $g^{0}$ to a player $i$, the strategy of this player may drift away from (1), which then will trigger strategy adjustments from the other players. Facing these induced deviations, if it is the best response for player $i$ to deviate further away from (1) then the Nash equilibrium is unstable. To investigate this type of local stability one often needs to specify how individual players update their action choices. A comprehensive review on population dynamics of strategic interactions is presented in [21].

In the next subsection we introduce another type of stability criterion which is independent of the particular microscopic competition dynamics.

2.2 Evolutionary stability

Stability of a mixed strategy can also be defined under the perspective of mutation and selection [17, 18, 19]. Let us consider a population of $N$ agents interacting with a mixed strategy, say $\vec{w}$. Suppose now a mutation occurs to a subpopulation of $n<N$ agents such that these $n$ agents adopt a different strategy (say $\vec{w^{\prime}}$). For this hybrid system, if the expected payoff $g$ of an agent in the unperturbed subpopulation is higher than the expected payoff $g^{\prime}$ of an agent in the mutated subpopulation, then the original strategy $\vec{w}$ is regarded as an evolutionarily stable strategy, otherwise it is an evolutionarily unstable strategy [22]. The concept of evolutionary stability is basic to evolutionary game theory and is very useful for understanding biological evolution [18, 23].

For the iterated RPS game, let us assume the mutated strategy of $n$ members is $(w_{r}^{\prime},w_{p}^{\prime},w_{s}^{\prime})$ while the remaining $(N-n)$ members adopt the NE mixed strategy $(w_{r},w_{p},w_{s})=(1/3,1/3,1/3)$, see Figure 2. At each game round every player competes with a randomly chosen member of the whole population. The expected payoff for a player in the unperturbed subpopulation is simply $g=g^{0}$, while that of a player in the mutated subpopulation is

[TABLE]

If $a>2$ we have $g>g^{\prime}$, players adopting the NE strategy have a higher expected payoff than players adopting the mutated strategy. By natural selection the mutated subpopulation should shrink in size ($n\rightarrow 0$), making the NE strategy evolutionarily stable. On the other hand, if $1<a<2$ we have $g<g^{\prime}$, then the NE strategy is not evolutionarily stable and can not persist under strategy mutations. In this latter parameter region the system actually has no evolutionarily stable strategy.

Why is the strategy $(1/3,1/3,1/3)$ evolutionarily stable when $a>2$ but unstable when $a<2$? The reason lies in the interactions within the mutated subpopulation. Consider two players adopting the mutated strategy. If they meet, the probability of tie is

[TABLE]

which is larger than $1/3$. As the mean payoff $a/2$ of a win-lose output is less than that of a tie output only for $a<2$, the mutated strategy is beneficial only in this parameter region.

In the above discussions, a mixed strategy $(w_{r},w_{p},w_{s})$ is defined at the level of individual players. We can also define a mixed strategy at the population level. In this latter perspective each individual may hold a fixed action (a pure strategy), then the mixed strategy describes the fractions of individuals adopting the different actions, which is more appropriate for studying strategic interactions in some biological systems. The same analysis of evolutionary stability can be carried out at the population level to answer the question of stable population composition under natural selection [18, 23].

2.3 Cooperation-trap strategies

If an iterated RPS system is in the mixed-strategy Nash equilibrium, every player stays in a safe position free of being exploited by the others. This is of course fine in terms of risk avoidance, but on the other hand as all players are interacting non-cooperatively they might miss the opportunity of achieving higher accumulated payoffs. Is it possible to sustain high degree of cooperation in this intrinsically non-cooperative game and beat the Nash equilibrium? For the two-person iterated RPS, it was demonstrated in [15] that there do exist simple strategies that are maximally fair and also maximally profitable to both players. Such strategies are referred to as cooperation-trap strategies [15] as they can induce an opponent player into complete cooperation. Here we offer an implementation of cooperation-trap strategies that improves the original protocol suggested in [15].

When the reward parameter $a>2$ the mean payoff to a player from a win-lose is higher than the payoff of a tie. An intelligent and rational player (say X) therefore has incentive to search for a strategy that maximizes the chance of wins while minimizes the chance of ties. The recipe of a simple cooperation-trap strategy goes as follow:

By default, player X acts in the cooperation mode. In this default mode, X avoids using one of the three actions (say $P$) and adopts the remaining two actions ($R$ and $S$) with equal probability $1/2$ in each game round. If player Y cooperatively responds to X by adopting action $P$, then both players get an equal expected payoff $a/2$ per round, which is a fair result and is higher than the value $g^{0}$ of the NE mixed strategy. 2. 2.

However if player Y exploits the cooperation mode of X by adopting action $R$, which returns an even higher expected payoff $(1+a)/2$ to Y, player X switches to the punish mode in the next $m$ game rounds. In this punish mode X employs the NE mixed strategy and adopts actions $R$, $P$ and $S$ with equal probability $1/3$. The expected payoff per round is then reduced to $g^{0}$ for both players. Player X is forgivable and she switches back to the default cooperation mode after each punish mode of length $m$.

If player X employs this cooperation-trap strategy, it is beneficial for player Y to abandon action $S$; furthermore if the punish mode duration $m$ is equal to or larger than a minimum value $m^{*}\equiv\lceil 3/(a-2)\rceil$, then it is optimal for Y to fix his action to $P$ in every game round. When $a>5$ we have $m^{*}=1$, suggesting that a minimum punish level is enough to sustain complete cooperation (Figure 3A).

When $1<a<2$ the mean payoff of a win-lose is less than the tie payoff. In this region the default cooperation mode of player X is modified to promote tie rather than win-lose. By default player X adopts the same action (say $P$) in every game round as long as player Y responds also with action $P$. If Y exploits X by adopting action $S$ in one game round, X switches to the punish mode in the next $m$ rounds and then switches back to the cooperation mode. It is also easy to verify that if $m\geq m^{*}\equiv\lceil 3(a-1)/(2-a)\rceil$ complete cooperation between the two players can be achieved (Figure 3A).

To achieve high degree of cooperation we need a proactive player X to initialize cooperation. Does the opponent player Y also need to be sufficiently intelligent to figure out the intention of X? This may not be necessary. Let us consider a totally myopic player Y who employs a mixed strategy $(w_{r},w_{p},w_{s})$ and changes this strategy in time to maximize his gain. Since action $R$ is better than $S$ when facing a player X adopting the cooperation-trap strategy, $w_{s}$ will evolve to zero and then the mixed strategy of Y becomes $(1-w_{p},w_{p},0)$. Player Y’s expected payoff $g^{(y)}$ per game round is compared with the corresponding $g^{(x)}$ of player X and the NE payoff $g^{0}$ in Figure 3B for the case of $a=5$ and $m=6$. We notice that $g^{(y)}$ is a strictly increasing function of $w_{p}$, indicating that complete cooperation ($w_{p}=1$) is the only fixed point of any gradual learning process of player Y.

Figure 3B also demonstrates that the expected payoff of player X is less than that of Y unless complete cooperation is reached. Indeed $g^{(x)}$ is even less than the NE payoff $g^{0}$ if player Y is not sufficiently cooperative. Player X may overcome such a drawback by further refinements of the cooperation-trap strategy. This issue and empirical evaluations of the cooperation-trap strategies will be further studied in a systematic way.

3 Conditional Response Patterns in Human Subjects

A convenient assumption of the preceding section is that players of infinite rationality (who must have excellent random-number generators!) make decisions based on certain mixed strategy $(w_{r},w_{p},w_{s})$ and modify this strategy according to feedback information from the iterated RPS. However such an assumption is often not realistic in competition processes involving human subjects. A person may not be well conscious of a mixed strategy, rather the decision-making is heuristic and is easily disturbed by environmental and psychological factors (e.g., automatic imitation of opponent’s actions [24]). Even if the human brain may have the capacity of implementing a mixed strategy, generating a random sequence of actions following this strategy is itself very demanding (the human brain performs poorly in randomization tasks [25, 26]).

How do people choose actions in the iterated RPS? As a first step to answer this difficult question, empirical investigations were carried out in the last few years by several research teams [27, 28, 29, 20]. Two of these experiments [27, 29] employed the all-to-all protocol with every person playing simultaneously against all the other players in the population, while [28, 20] employed the more traditional random pairwise-matching protocol, with every person only competing with another single player. Here we focus on the results of the latter as the experimental setting mimics decision-making under uncertainty.

A total number of $360$ university students participated in the experiment of Wang and Xu [28, 20]. These human subjects form $60$ groups with each group (population) containing six people. As shown in Figure 4A, at each game round $t=1,2,\ldots,300$, the six players of each population are randomly paired and they play once with their pair opponent (which might be different in different rounds) under the payoff rule of Figure 1B, and then every player receives feedback information about her/his own payoff in this round, her/his accumulated payoff, and the opponent’s action in this round.

In the experiment each population plays with a fixed reward parameter $a$ whose value ranging from $a=1.1$ to $a=100$. The empirical results demonstrated that the precise value of $a$ does not affect the qualitative dynamical behaviour of this finite-population system [20, 30].

3.1 Individual inertia effect and collective cycling

During the $300$-round iteration, each player leaves an action sequence $(s_{1}$, $s_{2}$, $\ldots$, $s_{300})$ with $s_{t}\in\{R,P,S\}$ being the action at the $t$-th round. We then get the preference vector of this player to the three actions, $(f_{r},f_{p},f_{s})$, where $f_{r}$ is simply the fraction of rounds the action being $R$ (similarly for $f_{p}$ and $f_{s}$, and $f_{r}+f_{p}+f_{s}\equiv 1$). We find that, consistent with the Nash equilibrium theory, the vectors $(f_{r},f_{p},f_{s})$ of the players are close to the mixed strategy $(1/3,1/3,1/3)$. For example $f_{r}=0.36\pm 0.07$, $f_{p}=0.32\pm 0.07$ and $f_{s}=0.32\pm 0.06$ at $a=2.0$ (mean and standard deviation, obtained by averaging over $72$ subjects) [20]. This observation is of course most natural given the rotational symmetry among the three actions.

The Nash equilibrium theory also predicts the choices of a player at two consecutive game rounds are completely independent of each other. But this is not what actually happened in the game. Instead there is considerable degree of temporal correlation within each action sequence. Especially, as demonstrated in Figure 5A, a player in each round is more likely to repeat the action of the previous round than to shift action either in the counter-clockwise ($R\rightarrow P$ or $P\rightarrow S$ or $S\rightarrow R$) direction or in the clockwise ($R\rightarrow S$ or $P\rightarrow R$ or $S\rightarrow P$) direction. This inertia effect is strongest at $a=1.1$, it weakens slightly with the increase of $a$ and is still very significant even at $a=100$ [20]. On the other hand, if a player does make a change, this change is symmetric in the sense that the probability of shifting action in the counter-clockwise direction is almost the same as that of shifting in the clockwise direction.

Although individual action changes from any action do not have directional preference, the empirical data reveal persistent counter-clockwise cycling in the collective behaviour of the population. The population’s collective state at each game round $t$ can be described by the vector $(n_{r},n_{p},n_{s})$ with $n_{r}$ being the total number of players adopting action $R$ (similarly for $n_{p}$ and $n_{s}$). The evolution of this so-called social state with $t$ then draws a trajectory in the social-state plane (Figure 4B), which is highly stochastic. To detect directional motions that signify deviation from equilibrium, a rotational angle $\theta$ is assigned to each social-state transition [20]. If the transition is associated with a counter-clockwise rotation around the centroid of the social-state plane, then $\theta$ is positive; if it is associated with a clockwise rotation around the centroid, then $\theta$ is negative; in all the other cases $\theta=0$ (Figure 4B). The net number of turns $C_{1,t}$ a trajectory cycles around the centroid of the social-state plane from the first to the $t$-th game round can then be obtained by adding up the rotation angles. As illustrated in Figure 5B for $a=2.0$, the accumulated cycle number $C_{1,t}$ has a linearly increasing trend with $t$, revealing persistent collective cycling along the $R\rightarrow P\rightarrow S\rightarrow R$ direction. The mean cycling frequency is $\nu\approx+0.03$ (one counter-clockwise cycle in about $35$ game rounds), and this value does not change significantly with the payoff parameter $a$ [20].

3.2 The conditional-response mechanism

The existence of weak but persistent collective cycling (Figure 5B) is apparently conflicting with the absence of directionality in the action shift behaviour of individual players (Figure 5A). Why a seemingly symmetric dynamics at the level of individual players results in asymmetric motion at the level of the whole population?

This apparent contradiction has been resolved in [20] by the key observation that players have different degrees of willingness to make a change, depending on whether the previous play was a win, a tie, or a lose. Consider the tendencies $W_{0}$, $T_{0}$ and $L_{0}$ of a person to choose the same action in two consecutive rounds if the earlier round is a win (W), a tie (T) or a lose (L), respectively. The empirical data reveal $W_{0}>T_{0}>L_{0}$, that is, the tendency of repeating an action increases with the payoff of the earlier round (Figure 5C). Such a microscopic pattern is qualitatively similar to the ‘win-stay lose-shift’ strategy of playing the iterated Prisoner’s Dilemma game, whose effectiveness has been confirmed by theoretical calculations [31, 32] and extensive computer simulations [33, 34]. The tendencies ($W_{-}$, $T_{-}$ and $L_{-}$) of shifting action in the clockwise direction and the tendencies ($W_{+}$, $T_{+}$ and $L_{+}$) of shifting in the counter-clockwise direction also depend considerably on the output of the previous round [20].

Inspired by these empirical conditional response patterns, a simple model of decision-making was proposed in [20] to understand the interactions in the iterated RPS. This model assumes the action shifting probabilities in each game round depend only on the game output in the previous round. For example if a player wins a play, then in the next round she/he has probability $W_{0}$ to stick to the same action, probability $W_{-}$ to shift action in the clockwise direction, and probability $W_{+}$ to shift action in the counter-clockwise direction. This conditional-response microscopic rule is characterized by six independent parameters $(W_{-},W_{+},T_{-},T_{+},L_{-},L_{+})$ which can be empirically fixed and it completely ignores all the possible higher-order complications in the decision process. Yet very encouragingly, this simple model quantitatively reproduce all the major experimental results, including the cycling frequencies and inertia effects measured in the $60$ different populations (for more details see [20]).

It is really interesting and unexpected that complicated decision-making processes of human subjects are statistically described by such a simple conditional-response mechanism. However we should also point out that the conditional-response model is at present only a phenomenological model. A missing link of basic significance is how the six independent conditional response parameters evolve in time as a result of learning. It could be that the conditional-response mechanism is just the result of certain lower-level learning dynamics. These questions need to be further studied from the neurobiology side [35, 36].

A counter-intuitive theoretical prediction of the conditional response model is that social-state cycling does not require microscopic asymmetry in conditional responses: it persists even if $W_{-}=W_{+}$, $T_{-}=T_{+}$ and $L_{-}=L_{+}$ provided that $W_{0}$, $T_{0}$ and $L_{0}$ are not all equal to each other [20]. This prediction has been verified by computer simulations, confirming that social cycling is indeed an emergent phenomenon.

3.3 Discussion on the effect of population size $N$

Under the random pairwise-matching protocol, each player has probability $1/(N-1)$ of encountering the same opponent in two consecutive game rounds. This probability decreases quickly with population size $N$, and consequently the action choices of different players are less and less correlated as $N$ increases. Since active decision-making is a costly mental process, as the action correlation decreases with $N$, the incentive for a player to change action should be weaker and weaker, especially if the player wins a previous round. Therefore we expect that the players will be less active in a larger population and the inertia effect of decision-making will be stronger; furthermore the win-stay probability $W_{0}$ will also increase with $N$.

The $N=2$ iterated RPS is special. As both players know about the opponent’s action history, it is a complete-information system. To avoid being the loser in this two-person game, both players should be very active and make their action choices most difficult to predict. We therefore anticipate that, (i) the inertia effect of individual players will be most weak, and (ii) the conditional response probabilities will only weakly deviate from $1/3$ and the players may not prefer to repeat a winning action.

The effect of population size will be systematically explored by laboratory experiments. These empirical studies may inspire the construction of refined learning models for the iterated RPS.

4 Rock-Paper-Scissors in Species Competition

Cyclic dominance is a ubiquitous phenomenon in ecological systems. For example the European honeybees invaded the local honeybees after being introduced to Japan but they were unprepared for the attacks from Japanese hornets, while the Japanese honeybees have developed a collective thermal defense mechanism against the hornets as a result of coevolutionary adaptation [37]. Another more quantitative example is the color polymorphism of male side-blotched lizards [38]. Field measurements revealed the frequencies of the three types of cyclicly dominant male lizards oscillate with a period of approximately six years and with a mutual phase shift of about two years [38]. Ecologists believe cyclic dominance to be a key factor contributing to ecosystem complexity, and they take the RPS game as a basic model of species competition and coexistence [39, 40, 7]. In this section we briefly review two simple microscopic ecological processes, the collision dynamics and the replicator dynamics.

4.1 The collision dynamics

We may consider an ecological system formed by three different species $R$, $P$, and $S$. The total number of individuals in the system is a fixed integer, $N$. Among these individuals $N\rho_{r}$ belong to species $R$, $N\rho_{p}$ to species $P$, and the remaining $N\rho_{s}$ to species $S$ ($\rho_{r}+\rho_{p}+\rho_{s}\equiv 1$). At each time interval $\delta t=2/N$ there is a competition between a randomly chosen pair of neighbouring individuals. If this competition is between two individuals of the same species, then both individuals survive. If it is between an individual of species $R$ and one of species $P$, then with probability $w_{p\leftarrow r}$ the $R$-individual is displaced by an offspring of the $P$-individual. The parameter $w_{p\leftarrow r}$ quantifies the dominance degree of the $P$ species to the $R$ species. Similarly the other two dominance degrees are denoted as $w_{r\leftarrow s}$ and $w_{s\leftarrow p}$, respectively. Notice that the expansion of one species is associated with the shrink of another species (constant-sum game).

Such a simple collision model and its various extensions have been investigated by many authors (see, e.g., reviews [39, 5, 7]). In the case of a well-mixed population each individual has equal chance of encountering any another one, then for $N\rightarrow\infty$ the species frequencies $\rho_{r}(t)$, $\rho_{p}(t)$, $\rho_{s}(t)$ as functions of time $t$ are governed by [41, 42]

[TABLE]

Equation (4), as a simple theory of cyclic dominance among three species, can be regarded as an extension of the celibrated Lotka-Volterra equation on the nonlinear predator-prey interactions between two species [43]. This evolution equation has a fixed-point solution of

[TABLE]

Notice that lowering the dominance degree $w_{r\leftarrow s}$ of species $R$ has an enhancing effect on its fixed-point frequency $\rho_{r}^{*}$. This is because decreasing $w_{r\leftarrow s}$ has the indirect consequence of suppressing the growth of species $P$ which preys on species $R$.

If the initial frequencies of the three species deviate from the fixed point $(\rho_{r}^{*},\rho_{p}^{*},\rho_{s}^{*})$, the system will not evolve into it but will move around it along a periodic orbit, see Figure 6. Equation (4) has infinitely many such limiting-cycle solutions, each of which is characterized by an invariant $C\equiv(\rho_{r}/\rho_{r}^{*})^{\rho_{r}^{*}}(\rho_{p}/\rho_{p}^{*})^{\rho_{p}^{*}}(\rho_{s}/\rho_{s}^{*})^{\rho_{s}^{*}}$ [42, 41]. $C=1$ at the fixed point $(\rho_{r}^{*},\rho_{p}^{*},\rho_{s}^{*})$, and it decreases continuously to zero as the distance of the limiting cycle to the fixed point increases.

Since the population state evolves along a limiting cycle, such an infinite ecosystem is only marginally stable. This has a very significant consequence for a finite ecosystem, namely species extinction is an unavoidable fate. Figure 6 illustrates a single evolutionary trajectory of a population with $N=1000$ individuals. Starting from an initial state $\rho_{r}=\rho_{p}=\rho_{s}=1/3$, the population state initially follows the periodic orbit, it then deviates more and more from this orbit due to the intrinsic stochasticity of the collision process. At the later stage the oscillatory magnitudes of the species frequencies become more and more pronounced and eventually only one species remains in the population. The final surviving species depends on the whole evolution process, but the species with the lowest value of dominance degree has the highest chance of survival (“survival of the weakest” [42]).

This prediction of species extinction does not agree with empirical observations. Different species do coexist in various real-world ecosystems containing only a finite number of individuals. There have been a lot of theoretical and experimental studies on this issue. A central aspect of real-world systems is its spatial structure [44]. Different species cluster into different local regions in a two- or three-dimensional space, and the interactions between different species occur only at the boundaries between these regions. Simulation results and theoretical computations reveal that all the species in such a spatial system have a high probability of survival even after an infinite evolution time. The species form very interesting entangled patterns, each region of a species is a spiral shape, and the boundaries of the regions move in time [45, 46].

In real-world ecosystems, individual animals also move actively in space. Increased mobility entails increased interactions with the other species and makes the population be more mixed. An interesting theoretical observation of [47] is that there is a sharp ‘phase transition’ of biodiversity in the Rock-Paper-Scissors ecosystem: if the species mobility exceeds certain critical value, the probability of species coexistence drops from $\approx 1$ to $\approx 0$ for a sufficiently large system. Therefore the RPS ecosystem can tolerate certain degree of species mobility but not too much.

4.2 The replicator dynamics

In many ecological systems, species conflict is not caused by the direct predator-prey interactions but is the result of competing for the same resources. For example, different strains of budding yeast cells may grow and reproduce in the same environment. One mother cell divides into two daughter cells after its body size exceeds certain critical value, so the number of yeast cells proliferates in an exponential manner under good nutrient conditions. If one strain of yeast cells has a higher growth rate than another strain, it will have a higher reproduction rate and its population size will then increase exponentially faster than that of the rival strain [45, 48].

The interactions among three types of yeast cells were studied in the experimental system of [45], and cyclic dominance in growth-rate advantage was observed. Similar cyclic dominance phenomena also exist in other engineered or naturally-occurring microbial ecosystems. For such an ecosystem, the reproduction rate of a particular species then depends strongly on the relative abundances of all the species. As a simple model we may again consider an ecosystem formed by three species $R$, $P$, and $S$. Let us assume that the reproduction rate $g_{r}$, $g_{p}$ and $g_{s}$ of the three species depend linearly on the species frequencies $\rho_{r}$, $\rho_{p}$ and $\rho_{s}$:

[TABLE]

where the parameter $c_{0}$ denotes the null reproduction rate when species competition is absent (for simplicity we assume it is the same for all the three species), which in general depends on the total number of individuals in the population. Notice that increasing the frequency $\rho_{p}$ of the $P$ species has a negative effect on the reproduction rate of the $R$ species, while increasing $\rho_{s}$ of the $S$ species has a positive effect on $g_{r}$ when $a>1$.

The expected number $n_{r}(t)$ of $R$-individuals evolves with time $t$ according to ${\rm d}n_{r}(t)/{\rm d}t=n_{r}(t)g_{r}$ and similarly for the other two values $n_{p}(t)$ and $n_{s}(t)$. The total population size $N(t)\equiv n_{r}(t)+n_{p}(t)+n_{s}(t)$ then evolves according to ${\rm d}N(t)/{\rm d}t=N(t)\overline{g}$, with the mean reproduction rate being $\overline{g}=\rho_{r}g_{r}+\rho_{p}g_{p}+\rho_{s}g_{s}$. If $\overline{g}(t)$ is always positive the total population will diverge with time. Since $\rho_{r}(t)\equiv n_{r}(t)/N(t)$ and similarly for $\rho_{p}(t)$ and $\rho_{s}(t)$, we have

[TABLE]

Equation (7) is independent of the null reproduction rate $c_{0}$. Such an evolution dynamics is usually referred to as the replicator dynamics [22, 19].

The deterministic dynamics (7) is easy to solve numerically. This equation has a fixed-point solution of $\rho_{r}=\rho_{p}=\rho_{s}=1/3$. If $a>2$, the population state $(\rho_{r},\rho_{p},\rho_{s})$ evolves towards the fixed point $(1/3,1/3,1/3)$ starting from any initial condition satisfying $\rho_{r}\rho_{p}\rho_{s}>0$, therefore this fixed point is a globally stable state and coexistence of all the three species is stable towards perturbations (Figure 7, top panel). At $a=2$, however, the evolution converges to a limiting cycle (middle panel of Figure 7). Similar to the collision model of the preceding subsection, the ecosystem is then only marginally stable and will eventually goes to species extinction. To remain biodiversity we need to consider again the spatial structure of species competition.

When $1<a<2$, the evolution does not converge to a fixed point, nor to a limiting cycle, but keep cycling with longer and longer period. The last phenomenon of non-periodic oscillation was first discussed in [49]. As shown in the bottom right panel of Figure 7, the frequency of each species (say $\rho_{r}$) jumps back and forth between the nearly extinct state ($\rho_{r}\approx 0$) and the overwhelmingly occupied state ($\rho_{r}\approx 1$). The jumps between these two situations occur very quickly while the residing time in each state become longer and longer. The reason that a species can recover from nearly extinction to predominance is mainly due to the assumption of offspring’s exponential proliferation. In the actual situation of a finite population, two of the species will be extinct inevitably.

The replicator dynamics has been extensively used in game theoretical studies [4]. It is a quantitative model system to discuss evolutionarily stable strategies [17, 18]. This dynamics can also be interpreted as a simple model of social learning through imitation [19].

5 Rock-Paper-Scissors in Market Price Competition

It is common to observe that different shops in a market sell the same item at different prices, and furthermore, the price of each shop is not a constant but changes non-monotonically in time (see, e.g., [50, 51]). Edgeworth first predicted the existence of persistent price cycles [52], arguing that under strong competition shops will lower their prices in small steps to attract more customers, but if the price hits a bottom level they will suddenly lift price to a much higher level. The RPS game has been used in the theoretical economics field to qualitatively describe price dynamics [53].

Suppose a simplified situation that each day there are $N$ new lazy buyers and $N$ new diligent buyers looking for a particular item. A lazy buyer enters into the first shop he found and buys the item, while a diligent buyer examines all the shops in the market and buys the item at the shop offering the lowest price. Let us further assume that there are only two shops selling this item and these two shops can choose to sell at three different price levels, the high price $h$, the medium price $m$, and the low price $l$.

Initially both shops may sell the item at the high price $h$, and the expected profit or payoff for each shop is $Nh$. If one shop (say X) now lowers the price to $m$, its expected payoff changes to $\frac{3N}{2}m$, which is higher than $Nh$ if $m>\frac{2}{3}h$, but the expected payoff for the other shop (Y) is reduced to $\frac{N}{2}h$. If $l>\frac{2}{3}m$, the best response of shop Y is to shift price from $h$ to the low value $l$, which will lead to an expected payoff of $\frac{3N}{2}l$ for itself and a reduced value $\frac{N}{2}m$ for shop X. But if $h>2l$, shop X will again respond by shifting to the high price $h$ which increases its expected payoff to $\frac{N}{2}h$, …, causing persistent price oscillations [54]. The above analysis demonstrates that in the parameter range of $\frac{2}{3}h<m<\frac{3}{2}l$ and $2l<h<\frac{9}{4}l$, there is the cyclic-dominance of the medium price $m$ beating the high price $h$, the low price $l$ beating the medium price $m$, and the high price $h$ beating the low price $l$.

If this simple market system stays in the mixed-strategy Nash equilibrium, the mixed strategy $(w_{l}^{0},w_{m}^{0},w_{h}^{0})$ for each shop to choose the low, middle, and high price is

[TABLE]

Notice that $w_{l}^{0}>w_{m}^{0}>w_{h}^{0}$, meaning that each shop should choose the low price with the highest probability and the high price with the lowest probability. For the representative parameter set $l=1$, $m=1.45$ and $h=2.1$, we have $\rho_{l}^{0}\approx 0.457$, $\rho_{m}^{0}\approx 0.332$ and $\rho_{h}^{0}\approx 0.211$. The expected payoff per day (normalized by $N$) is $g^{0}=mhl/[ml+(m-l)h]$, which is always larger than $(6/5)l$ but less than $(4/3)l$.

But in reality the prices of different shops are highly entangled [50, 51]. For our two-shop toy model, a simple microscopic process to mimic correlated decision-making is the noisy best-response dynamics with a single parameter $\beta$ [55]. Knowing the price $s_{t-1}^{(x)}$ of shop X at day $(t-1)$ but uncertain of its price at the next day $t$, shop Y chooses its price $s_{t}^{(y)}\in\{l,m,h\}$ at day $t$ according to the conditional probability distribution

[TABLE]

where $g(s^{(y)}|s^{(x)})$ is shop Y’s payoff (normalized by $N$) in state $s^{(y)}$ while the other shop X chooses price $s^{(x)}$, e.g., $g(l|l)=l$, $g(m|l)=m/2$, and $g(h|l)=h/2$.

The steady-state marginal probabilities ($w_{l}$, $w_{m}$, and $w_{h}$) of choosing the three prices are easy to compute if both shops are governed by the same stochastic dynamics (9). This steady-state distribution is very different from the NE mixed strategy (see Figure 8, left). If $\beta=0$, there is no selection, $w_{l}=w_{m}=w_{h}=1/3$, and the expected payoff per day is $g=(2h+3m+4l)/9$. As $\beta$ increases the expected payoff changes in a non-monotonic way, and there is an optimal value of $\beta$ at which the expected payoff reaches the global maximum value much higher than the NE value $g^{0}$ (Figure 8, middle).

We can study the collective behaviour of this two-shop toy model following the same method of Section 3.1. When $\beta>0$ there is persistent cycling in the system (Figure 8, right), which is consistent with the Edgeworth price cycle [53, 52]. The cycling frequency is an increasing function of $\beta$ (positive cycling direction is high price $\rightarrow$ middle price $\rightarrow$ low price $\rightarrow$ high price).

Real-world markets are of course much more complex than the model discussed here. The key points we want to emphasize are (i) cyclic dominance among different price levels do occur in real markets [50, 51], and (ii) it is much more beneficial to make intuition-guided decisions rather than to follow the Nash-equilibrium mixed strategy in price competition. The second point is indeed closely related to the debated issue of rationality (the reader may consult [56] for more discussions).

6 Outlook

The Rock-Paper-Scissors game is a simple game that helps improving our understanding on many complex competition issues (species divergence, price cycling, human decision-making, rationality and cooperation and so on). This game is the simplest model system for studying the non-equilibrium statistical mechanics of non-cooperative strategic interactions, and it can serve as a starting point to enter into the interdisciplinary field between statistical physics and game theory.

In this brief review we have left untouched the issue of possible phase transitions. If the Rock-Paper-Scissors game is played on an infinite lattice or a complex network, are there competition-driven critical phenomena and how to quantitatively describe these behaviours? At present, how players adjust their decision-making parameters is still quite unclear and is largely ignored in theoretical investigations. A lot of empirical and theoretical efforts are needed in these directions.

Acknowledgement

I thank Prof. Bin Xu and Dr. Zhijian Wang for collaborations, and thank Prof. Bin Xu for a critical reading of the manuscript. This work is partially supported by the National Basic Research Program of China (grant number 2013CB932804) and by the National Natural Science Foundations of China (grant numbers 11121403 and 11225526).

Hai-Jun Zhou received his Doctorate from the Chinese Academy of Sciences in 2000. His dissertation was on structural transitions in biopolymers. Since 2005 he has been a research professor at the Institute of Theoretical Physics of the Chinese Academy of Sciences. Currently he is working on statistical physics and its interdisciplinary applications, including spin glasses and combinatorial optimization, percolation processes in complex networks, and collective dynamics in game systems.