Learning collaborative team behavior from observation
Expert Systems with Applications: An International Journal
| Hi-index | 0.00 |
This paper proposes and explores an interaction protocol for incorporating helpful behavior into agent teamwork. In the proposed Mutual Assistance Protocol (MAP), an agent can directly assist a teammate who requests help, provided that the two agents jointly determine, based on their individual beliefs, that the expected outcome of the help act is in the interest of the team. This distributed decision is reached through a bidding sequence similar to the one in the Contract Net Protocol. The deliberation about help is approximate in that each agent only assesses the team impact of the change to its own individual plan. The paper introduces two versions of the protocol: Action MAP, in which the helper performs an action within a teammate's individual plan, and Resource MAP, in which one or more helpers provide resources to a teammate. Both versions include refinements for the handling of simultaneous help requests. A cooperative game simulation demonstrates the advantages of Action MAP over action help protocols that use unilateral decision mechanisms, and over teamwork scenarios without help. The experiments show how the team performance depends on: the teammates' mutual awareness of each other's abilities, dynamic disturbance in the environment, communication costs, and computation costs.