Should I Trust my Teammates? An experiment in Heuristic Multiagent Reinforcement Learning

摘要

Trust and reputation are concepts that have been
\ntraditionally studied in domains such as electronic
\nmarkets, e-commerce, game theory and bibliometrics,
\namong others. More recently, researchers
\nstarted to investigate the benefits of using these
\nconcepts in multi-robot domains: when one robot
\nhas to decide if it should cooperate with another one
\nto accomplish a task, should the trust in the other be
\ntaken into account? This paper proposes the use of
\na trust model to define when one agent can take an
\naction that depends on other agents of his team. To
\nimplement this idea, a Heuristic Multiagent Reinforcement
\nLearning algorithm is modified to take
\ninto account the trust in the other agents, before selecting
\nan action that depends on them. Simulations
\nwere made in a robot soccer domain, which extends
\na very well known one proposed by Littman by expanding
\nits size, the number of agents and by using
\nheterogeneous agents. Based on the results it
\nis possible to show the performance of a team of
\nagents can be improved even when using very simple
\ntrust models.