Effective deadlock resolution with self-interested partially-rational agents

This paper pertains to distributed deadlock resolution in settings populated with self-interested bounded-rational autonomous agents. In particular it reports the results of extensive experimentation with 66 agents, each using a deadlock resolution strategy developed and maintained throughout the experiment by a different human decision maker. While from the game-theoretic perspective a simple equilibrium-based solution can be engineered for the problem, it is shown that such a solution fails to hold with bounded-rational agents, even when its principles are thoroughly explained to the individuals that maintain the agents’ strategies. Instead, we show that the system converges to a steady-state, in which the agents use a rich set of different strategies, varying in their performance, as each of them has a different belief regarding its ability to improve, based on the behaviors of the others. To improve system performance, we develop and implement a restructuring heuristic that changes the input each agent receives, as a means of affecting the agents’ decisions to better align with the desired solution. The restructured deadlock presented to each agent is based on former deadlocks it had experienced. Our experiments demonstrate the effectiveness of the restructuring heuristic in facilitating a new steady-state in which the system as a whole substantially improves its performance. The efficiency of the method, in terms of the size of the set of former experiences required for effectively restructuring the agents’ input, is demonstrated through a comparison with a neural network implementation of input restructuring, showing a substantial advantage to the restructuring heuristic.