Maximization of Potential Information Flow as a Universal Utility for Collective Behaviour

Daniel Polani, Department of Computer Science, University of Hertfordshire, United Kingdom

In the search for plausible mechanisms for the emergence of coordinated, mutually beneficial behaviour of agent collectives, the main challenge is to find local incentives for agents to coordinate and cooperate. The difficulty is that cooperation may incur short-term costs the long-term advantage of cooperation may not be immediately visible to the agents.

To move away from design-based to principle-based behaviour strategies, in recent research we studied empowerment, the potential information flow that an agent is able to induce in its environment through its perception-action loop. It was found that empowerment is a versatile and intrinsic quantity directly able to single out states of "interest" to the agent, based entirely on its perception-action loop and without reference to other, "external" value systems. In the present work, we study how even short-term empowerment maximization is sufficient to give rise to organized behaviour in agent collectives.