Cognitive Instance-Based Learning Agents in a Multi-Agent Congestion Game
Paul Scerri, David Reitter · 2013
Simultaneous learning by multiple agents can lead to undesirable and unwanted dynamics because the learning creates a non-stationary environment for the other agents to learn against. Although humans often face this challenge, humans can often converge to good cooperative solutions in reasonable amounts of time. In this paper, we empirically compare a multi-agent learning approach inspired by human ways of learning against a more numerically intensive agent way of learning. Specifically, agents must repeatedly traverse a graph and agents on the same edge will interfere with one another so learning is required to find uncongested routes. We find that human inspired instance-based learning performs at least as well as more quantitative and communication intensive approaches. Even when the overall system performance is closely matched, a look at the details of how the result was achieved shows considerable differences between approaches. Some of the advantage can be attributed to the instance-based learning’s preference for sticking with known good solutions, because this creates stability that allows the other agents to learn. However, external disturbances, changes to the underlying system, can be more problematic for instance-based learning precisely because so much history is used. 1.