Reconfiguration in a Dynamic Environment
5.1 Dealing with Dynamics
A dynamic environment requires flexible behavior to ensure successful teamwork. Even though the first stages of teamwork have been extensively discussed in the MAS and AI literature (Cohen et al., 1997; Nair et al., 2003; Pynadath and Tambe, 2002; Shehory, 2004; Shehory and Kraus, 1998), the resulting team action (or plan execution) has received relatively little attention. Let us analyze this phase now.
To maintain a collective intention during plan execution, it is vital that agents replan properly and efficiently in accordance with the circumstances. When some team members cannot realize their individual actions, or, on the positive side, some others are presented with new opportunities, re-planning takes place. This intelligent re-planning is the essence of the reconfiguration problem, discussed for the first time independently by Tambe (1996, 1997) and by Dunin-K ¸eplicz and Verbrugge (1996, 2001a).
During reconfiguration, adaptations of the original plan may be done from scratch, for the price of losing what has been achieved before. For resource-bounded agents, it is much more efficient to smartly adapt the previous results to the current situation. Such intelligent re-planning implies a natural evolution of the team’s commitment, including the evolution of plans and motivational attitudes involved. These changes are methodologically treated in a generic reconfiguration algorithm formulated in terms of the consecutive stages of teamwork and their complex interplay.
Teamwork in Multi-Agent Systems: A Formal Approach Barbara Dunin-K ¸eplicz and Rineke Verbrugge 2010 John Wiley & Sons, Ltd
82 Teamwork in Multi-Agent Systems
5.1.1 Collective Commitments in Changing Circumstances
Now that all teamwork attitudes have been intuitively and formally characterized in the static part of TEAMLOG (see Chapters 2, 3 and 4), they should be confronted with the paradigmatic situation justifying their creation, namely a complex, dynamic environment. After all, the proof of the pudding is in the eating.
To study teamwork and its dynamics, we will isolate and analyse separately the three essential aspects of team cooperation and coordination in a distributed environment. These are construction, maintenance and realization of the type of collective commitments that optimally fit the application domain, the group structure and the situation. This can be done by the system developer at design time or by the initiator at runtime. On the general issue of tuning collective commitments, see Chapter 4. Throughout this chapter, we will use the generic notion of collective commitment C-COMMG,P, abstracting from any particular type of commitment. As reconfiguration amounts to intelligent replanning, we will naturally focus on a team’s social plan which is an obligatory element of any group commitment. This way our approach to reconfiguration acquires universality, transcending commitment types.
5.1.2 Three Steps that Lead to Team Action
In many BDI systems, teamwork is modeled explicitly (Aldewereld et al., 2004; Grosz and Kraus, 1996; Levesque et al., 1990; Tambe, 1996, 1997; Wooldridge and Jennings, 1999). An explicit model helps the team to monitor its performance and to re-plan efficiently, in accordance with the circumstances, when team members fail to realize their actions or new opportunities appear. A commonly recognized model of cooperative problem solving (CPS) has been provided by Wooldridge and Jennings (1996, 1999). We adapted their four- stage model, containing the consecutive stages of potential recognition, team formation,
plan formation and team action for the sake of our analysis. However, especially with
respect to collective intentions and collective commitments, our approach differs from the one in Wooldridge and Jennings (1996, 1999).
As advocated above, we study teamwork starting from potential recognition, assuming for simplicity that there is an agent-initiator who knows the overall goalϕ and takes the
initiative to realize it. This initiator is responsible for potential for cooperation among agents available at the time. The next step is team formation, leading to a collective intention between members of a successfully created team. The subsequent stage of plan- ning, realized collectively in the most advanced case, results in the strongest motivational attitude, that is collective commitment. These complex preparations are finally concluded in team action.
An unpredictable and dynamic environment strongly influences teamwork, which becomes unpredictable to some extent when adjusting to actual circumstances. Therefore, modeling teamwork requires methods and techniques reflecting dynamics of its stages. Most of the time these methods originate from (Distributed) Artificial Intelligence; however, their specific variants have been created especially for multi-agent applications; see Durfee (2008), Jennings and Wooldridge (2000) and Wooldridge (2009) for extensive discussions.
Reconfiguration in a Dynamic Environment 83
In our reconfiguration story, we abstract from strictly technical aspects like methods and algorithms meant to realize stage-related procedures. Instead, our primary methodologi- cal goal is to characterize the stages of teamwork in a generic way, with an emphasis on their cooperative essence: the evolution of informational and motivational attitudes of team members. We focus on defining the final results of these stages in terms of agents’ motivational stance. Such an approach will be profitable in clarifying the nature of dependencies between the agents involved. For example, some of them do domain problem solving, while others are responsible for the proper organization of teamwork.
The rest of this chapter is structured as follows. Section 5.2 presents a detailed introduc- tion to the four stages of teamwork, including formal definitions of corresponding agent attitudes. Subsequently, the general ideas behind our reconfiguration method are explained and the reconfiguration algorithm is presented in the central Section 5.3. Finally, the algorithm is illustrated by an example application, extensively discussed in Section 5.4.