Emergent Cooperative Behavior in Multi-Agent Grid Resource Competition

This study investigates the strategic decision-making and cooperative behaviors of AI agents in a grid resource competition game. Agents learn to flexibly switch between attack, track, explore, and avoid strategies based on environmental factors. Remarkably, they exhibit altruistic rescue behavior, abandoning high-value resources to defend teammates.

Details

The study constructs a grid resource competition game where two teams of AI agents compete over resources. Each agent is controlled by a spiking neural network that selects among four high-level strategies: attack, track resources, explore randomly, and avoid enemies. Through evolutionary optimization, the agents learn to adaptively switch between these strategies based on environmental factors like resource energy, enemy proximity, self-energy, and hunger. The most remarkable finding is the emergence of altruistic rescue behavior - when a teammate is being flanked by enemies, an agent will voluntarily abandon its current high-value resource point to attack the enemies, preferentially targeting the lower-energy opponent. This cooperative behavior arises without any explicit reward for teamwork, suggesting it may be driven by social cognition mechanisms like social referencing and team identity based on communication signals. The study analyzes these phenomena from the perspectives of strategic decision-making, neural network bases, social functions of communication, and evolutionary conditions for altruism, with implications for understanding human cooperation.