This is the most serious negative case, generally leading to system-failure. Methods for checking whether failure of a certain action execution blocks the goal ϕ are again
88 Teamwork in Multi-Agent Systems
sensitive to the planning method and the application in question. Even though such checks might be viewed as an idealized notion, still they will be performed in the reconfiguration algorithm for efficiency reasons.
5.3.1 Continuity and Conservativity
When formulating the reconfiguration algorithm, we have chosen some intuitive properties corresponding to classical strategies adopted in backtracking. We postulate that the system behavior should preserve continuity . This means that, if an obstacle appears, the problems are solved by moving up in the hierarchy of teamwork stages, but as little as possible. The stages are ordered from potential recognition at the top to team action at the bottom. Thus, one moves to the nearest point up in the hierarchy of stages where a different choice is possible. If such a point does not exist anymore, the reconfiguration algorithm fails. In other words, depth-first search is used.
As regards generic stage-related procedures, a context-dependent question arises which ‘local’ results should be preferred. The answer calls for a domain-specific notion of the distance between teams, goals, plans, etc. At first glance, it seems that for a wide class of applications, it is justified that the system behaves in a conservative way (see also Nebel and Koehler, 1992). Importantly, conservativity (or inertia) entails that the collective commitment in question should change as little as necessary at every round of its evolution. The continuity criterion is application-independent and it determines the overall structure of the algorithm.
5.3.2 Reconfiguration Algorithm= Teamwork in Action
The reconfiguration algorithm presented below is meant to be generic: a pattern of behav- ior is described in terms of complex stage-associated procedures, called:
potential−recognition,team−formation,task−division,
means−end−analysis,action−allocation,plan−execution without fixing any particular method or strategy. Input and output parameters, as well as other conditions of these procedures, are commented upon in the algorithm below (see Algorithm 1). As the environment is dynamic and often unpredictable, each of these procedures may succeed or fail – in this sense all stages of teamwork have a similar struc- ture. Therefore, we use labels and appropriate GOTO statements for each stage-associated procedure to make the overall structure of the algorithm transparent. We introduce the predicate succ to denote that a procedure was performed successfully and the predicate
failed to denote that a procedure was performed but failed. Note that in the reconfiguration
algorithm, when the predicate failed occurs (say, in a line such as failed(division(ϕ))), then it does not cause any execution of the action that is its parameter (heredivision(ϕ)), it solely checks the status of the latest execution of this action; similarly for succ.
As indicated, the essence of the reconfiguration algorithm is the evolution of collective commitments and underlying plans. This inevitably leads to a revi-
sion of relevant individual, bilateral and collective motivational attitudes. In the
algorithm, phases of belief revision and motivational attitude revision are distin- guished, without further refinement, that is without splitting collective motivational
Reconfiguration in a Dynamic Environment 89
attitudes into individual and social ones. They are realized by abstract procedures: BeliefRevisionandMotivationalAttitudeRevision.
Pragmatically, the proper treatment of revision is ensured by the obligation that agents communicate about changes. On the other hand, the presence of social commitments and some form of awareness (for example common belief) about them solely between partners, together with the conservativity assumption, ensures that motivational attitudes revision is as efficient as possible.
The final outcome of the reconfiguration algorithm with respect to the overall goalϕ is
either system failure or success, realized by the generic proceduressystem−failure(ϕ)
and system−success(ϕ), respectively. It is also assumed that all the required infor-
mation is available at design-time.
In the algorithm (see Algorithm 1), the finite pool of agents T and H , denoting the
collection of possible teams from T , are global parameters. Even though both refer to
finite sets, it is possible that T and H evolve during the teamwork process; this makes
reconfiguration suitable for an open environment.
Since the reconfiguration algorithm is formulated in a generic way, it needs to be tailored for specific applications. Undoubtedly, whatever domain is considered, the stage-associated procedures, including belief revision and motivational attitude revision, remain complex. The attitude revision during reconfiguration will be carefully treated in Chapter 6. Importantly, the algorithm structure is based on backtracking search, leaving room for improvements like informed search methods, including varieties of hill-climbing (Foss and Onder, 2006; Koza et al., 2003) for particular applications.
5.3.3 Cycling through Reconfiguration
Reconfiguration is essentially a controlled type of evolution. To make this clear, let us focus on the failure points of the main stages of teamwork (see Algorithm 1):
1. The failure of potential recognition (see label A), meaning that agent a does not see
any potential for cooperation with respect to the goalϕ, leads to failure of the system.
2. The failure of team formation (see label B), meaning that the collective intention C-INTG(ϕ) cannot be established among any of the teams from H , requires a return to potential recognition to construct a new collection of potential teams.
3. The failure of task division (see label C) requires a return to team formation, in order to establish a collective intention in the chosen new team fromH . This may be viewed
as the reconfiguration of the team.
4. The failure of means-end analysis (see label D) requires a return to task division in order to create a new sequence of tasks, that would be the subject of a new round of means-end analysis.
5. The failure of action allocation (see label E) requires a return to means-end analysis in order to create a new sequence of actions that would be allocated to team members.
When, finally, a collective commitment is successfully established, the failure of some action executions from the social plan P leads to the evolution of the collective com-
mitment, as a result of conservative replanning. This evolution will be discussed in Chapter 6.
90 Teamwork in Multi-Agent Systems