Learned Collusion
This addresses the challenge of enabling cooperation in automated systems like AI agents, but it appears incremental as it builds on existing Q-learning frameworks with added biases.
The paper tackles the problem of fostering collusion or cooperation in multi-agent systems by introducing a family of automata based on Q-values with systematic biases, such as favoring cooperation, and finds that stable equilibrium biases can be learned under converging dynamics without tacit agreement, working across various payoff and monitoring structures regardless of initial Q-values.
Q-learning can be described as an all-purpose automaton that provides estimates (Q-values) of the continuation values associated with each available action and follows the naive policy of almost always choosing the action with highest Q-value. We consider a family of automata based on Q-values, whose policy may systematically favor some actions over others, for example through a bias that favors cooperation. We look for stable equilibrium biases, easily learned under converging logit/best-response dynamics over biases, not requiring any tacit agreement. These biases strongly foster collusion or cooperation across a rich array of payoff and monitoring structures, independently of initial Q-values.