Dynamically generated cyclic dominance in spatial prisoners dilemma games

We have studied the impact of time-dependent learning capacities of players in the framework of spatial prisoner’s dilemma game. In our model, this capacity of players may decrease or increase in time after strategy adoption according to a step-like function. We investigated both possibilities separately and observed significantly different mechanisms that form the stationary pattern of the system. The time decreasing learning activity helps cooperator domains to recover the possible intrude of defectors hence supports cooperation. In the other case the temporary restrained learning activity generates a cyclic dominance between defector and cooperator strategies, which helps to maintain the diversity of strategies via propagating waves. The results are robust and remain valid by changing payoff values, interaction graphs or functions characterizing time-dependence of learning activity. Our observations suggest that dynamically generated mechanisms may offer alternative ways to keep cooperators alive even at very larger temptation to defect.

💡 Research Summary

The paper investigates how time‑dependent learning capacities of agents affect the evolution of cooperation in spatial Prisoner’s Dilemma (PD) games. In the classic spatial PD, each player interacts with its neighbors on a lattice (or other network) and may adopt a neighbor’s strategy with a probability that depends on the payoff difference. The authors enrich this framework by allowing the “learning activity” – the propensity to imitate a neighbor – to change after a strategy adoption event. Two distinct temporal profiles are examined: (1) a decreasing learning activity, where an agent’s ability to learn drops immediately after it switches strategy and remains low for a prescribed period; and (2) a temporarily restrained learning activity, where learning is completely suppressed for a finite interval τ after adoption and then restored to its original level. Both profiles are implemented as step‑like functions, but the authors also test smoother functional forms (linear, exponential) to confirm robustness.

Using extensive Monte‑Carlo simulations on square lattices (von Neumann neighborhoods) as the primary substrate, the authors explore a wide range of payoff parameters (temptation T, sucker’s payoff S) and initial conditions. They also repeat the experiments on random Erdős–Rényi graphs and Watts–Strogatz small‑world networks to verify that the observed phenomena are not artifacts of a particular topology.

Key findings for the decreasing‑learning case: after a player adopts a new strategy, its reduced learning activity makes it less likely to be overwritten by a neighbor. Consequently, when defectors (D) infiltrate cooperative (C) domains, the newly converted cooperators are “protected” from being reconverted to defectors, allowing C clusters to recover and expand. This creates a feedback loop—“learning suppression → defensive reinforcement”—that markedly raises the stationary fraction of cooperators, even for high temptation values (T ≫ 1). The effect persists across network types and payoff settings, indicating a robust mechanism for cooperation enhancement.

Key findings for the temporarily restrained case: the enforced learning pause creates a period of strategic inertia. During this pause, neither cooperators nor defectors can spread, but once the pause ends, the agents’ learning activity resumes at full strength. This sudden release triggers rapid, wave‑like invasions. Defectors may surge forward, but cooperators, having reorganized during the pause, can counter‑attack, leading to a cyclic dominance pattern (D → C → D). The spatial manifestation is a series of propagating fronts, spirals, or other wave structures that maintain a dynamic coexistence of strategies. This cyclic dominance prevents any single strategy from monopolizing the system, thereby preserving diversity even under strong temptation.

Robustness checks confirm that (i) the qualitative behavior does not depend on the exact shape of the time‑dependence function; (ii) varying the length of the suppression interval τ modulates the wavelength and speed of the propagating fronts but does not eliminate the cyclic dominance; (iii) the phenomena survive on heterogeneous interaction graphs, indicating that the mechanisms are not limited to regular lattices.

The authors conclude that dynamically generated learning constraints can serve as alternative routes to sustain cooperation in hostile environments. The decreasing‑learning mechanism acts as a protective shield for cooperators, while the temporarily restrained mechanism introduces a self‑organizing cyclic dominance that keeps the system in a non‑absorbing, heterogeneous state. These insights suggest that real‑world systems—where individuals may temporarily lose the ability or willingness to imitate others due to fatigue, cultural taboos, or institutional rules—could exploit similar dynamics to maintain cooperative behavior despite high incentives to defect.

Future directions proposed include incorporating heterogeneous learning capacities across agents, coupling the learning dynamics to external environmental fluctuations, and extending the model to multi‑strategy games (e.g., inclusion of punishers, loners, or conditional cooperators). Such extensions would bring the theoretical framework closer to empirical observations in social, economic, and biological contexts, where time‑varying propensity to learn and imitate is a common feature.