Learning team coordination constraints through execution

Agents working together in teams can tackle user-defined tasks more complex than those they can perform as individuals. However constructing such teams remains a difficult challenge. In particular current approaches to designing agent teams are highly labor-intensive. Human designers must deal with overwhelming complexity in trying to manage the large number of interactions and dependencies that may exist between agent activities. Even if the designer is able to come up with a plan that seems to work, he cannot be sure that it will continue to work in all possible situations. We propose to use machine learning techniques to assist a user in building robust, multiagent team plans. This is done by logging information during team plan executions and attempting to find the cause of failure from this data. We present a method for learning temporal coordination constraints on actions in a multiagent reactive plan. We also briefly discuss the effect of new coordination constraints on team organization.