Advances in Artificial Intelligence

Modeling Role-Based Agent Team* (p. 12)
1 Introduction
Teamwork is becoming increasingly important in many dynamic multi-agent systems [13]. Agents in a team need to form joint mental states which drive agents to act together as a team and form the interactions leading their individual actions to team efforts [5, 7]. To simulate teamwork, a teamwork language is demanded to explicitly express the mental states underlying teamwork. In our opinion, the effective design of a teamwork language requires two aspects to consider. First, it should be able to handle unexpected uncertainties occurred in complex and dynamic domains, such as dynamic changes in team’s goals, team members’ unexpected failures to fulfill their responsibilities, decision-making in dynamic environment, and dynamically backing up other team members.

Second, considering the perspective of software engineering, the teamwork language would better allow specify teamwork knowledge conceptually for being reused, particularly, team plans are better specified in terms of abstract entities, instead of specific agents, so as to be reused by different teams of agents. A lower level of abstraction, role, is currently used by many researcher of multiagent systems to close this gap [14, 9, 11, 6, 16]. Biddle and Thomas’s role theory views role as the concept of partitioning behaviors and emphasizing coordination and cooperation [2].

Becht’s ROPE (Role Oriented Programming Environment for multiagent systems) uses roles to decouple the organization of agents from the structure of cooperation processes [1]. Cooperation process is designed from a global perspective and largely independent of concrete agents so that shifting cooperative behavior does not require changing agents (agents can fill multiple roles and switch between them). Stone and Veloso introduced roles as a mechanism for specifying an agent’s internal and external behaviors and decomposing team tasks [12], they then used this to model robot soccer. A formation decomposes the task space by defining a set of roles. There are as many roles as there are agents in the team, so that each role is filled by one agent. The mapping between agents and role is not pre-specified.

In this paper, we propose a role-based teamwork language. Different with the existing work, we use roles and role variables distinguish static (by roles) and dynamic (by role variables) action associations, and when delegating roles and role variables in a plan to agents, we have the agents form s joint mental state to enforce the execution of the plan as a team effort, particularly the sub-actions in the plan will be executed coherently. Our concepts of role and role variable enable our mechanisms of task decomposition and delegation, by which role-based plans drive agents to actually execute teamwork. Our mechanism of task decomposition is based on a notion of responsibility, which is defined in terms of what a responsibility contains and how a responsibility impacts the mental states of the agent(s) taking the responsibility. Our mechanism of task delegation has three steps:

1) a team task is translated to a team responsibility which is represented by a graph,

2) through decomposing the team responsibility graph to individual responsibility graphs, a team task is decomposed to individual sub-tasks, and

3) individual sub-tasks are delegated to agents.