Final Activity Report Summary - MASLOG (Formal methods for multi-agents systems)
The 'Formal methods for multi-agents systems' (MASLOG) project aimed at developing a logical framework for multi-agent systems, where agents represent autonomous component-based processes of computer systems. Agents are in essence autonomous entities making decisions and acting on the behalf of the user, preserving their interests while interacting with other entities in the environment. One of the major challenges in the multi-agent systems area concerns the understanding and the command of coalition and negotiation mechanisms between agents, as -in general- agents do not share the same interests.
The project has adopted an approach based on the paradigm of games -autonomous entities are modelled as players-, and explored different frameworks:
- First, the study has considered multi-player games where groups of players may form a coalition, against the rest, to enforce plays to be winning or favourable by fixing appropriate strategies. A highly expressive logical formalism was developed to specify strategies that are meant to be combined in coalition or that have a commitment feature. On top of this, any statement can be processed to compute the specified strategies. Moreover, the formalism is proved to subsume existing proposals, not only those mentioned in the original bibliography, but any other proposed so far. The study's work is published in the proceedings of the 5th International Symposium on Automated Technology for Verification and Analysis.
- Second, the complex setting of imperfect information games was investigated, where players only partially observe the moves of the others; hence they have uncertainty on the current configuration of the game. A mathematical theory was proposed, based on topology over the space of infinite sequences (representing the plays in the game), which enables one to reason on sets of observationally equivalent sequences (from the point of view of one player). Significant results for the one-player case, and applications to the diagnosis of discrete event systems have already been obtained; a publication in the 9th International Workshop on Discrete Event Systems is forthcoming.
- Third, optimality criteria on the coalition policies have recently been investigated, based on search techniques of the Artificial Intelligence (AI) field. For 'safety' games, i.e. games where the objective is to avoid 'bad' configurations, whether the observation capabilities of the players are sufficient for a coalition to win the game is in itself a challenging problem. Inspired from control theory, a sufficient condition was considered for the existence of a solution, which is decidable (i.e. which can be checked automatically). The research team studied a way to fix situations where this condition fails: by 'fixing', we mean to add communications between players, thereof augmenting their observation power to make the condition true. Penalising communication (by a cost function) led to solve optimisation problems by AI search techniques: the procedure renders an optimal communication addendum amidst players.
- Fourth, an interdisciplinary collaboration aims at bridging the gap between automated planning techniques and formal methods that are used for software verification. Automated planning is a field of AI. The basic problem consists in automatically choosing and organising a set of actions to reach a given objective in an optimal manner. Application areas include defence, business, and space. In software verification, the basic problem is also to determine whether a state of the software is reachable (typically an undesirable state where a deadlock or an overflow exists), or more generally whether all executions satisfy a given property. The close proximity between planning and verification is evident but has never been the subject of a solid theoretical study; these similarities are analysed as also the differences with an approach under game theory angle.
The project has adopted an approach based on the paradigm of games -autonomous entities are modelled as players-, and explored different frameworks:
- First, the study has considered multi-player games where groups of players may form a coalition, against the rest, to enforce plays to be winning or favourable by fixing appropriate strategies. A highly expressive logical formalism was developed to specify strategies that are meant to be combined in coalition or that have a commitment feature. On top of this, any statement can be processed to compute the specified strategies. Moreover, the formalism is proved to subsume existing proposals, not only those mentioned in the original bibliography, but any other proposed so far. The study's work is published in the proceedings of the 5th International Symposium on Automated Technology for Verification and Analysis.
- Second, the complex setting of imperfect information games was investigated, where players only partially observe the moves of the others; hence they have uncertainty on the current configuration of the game. A mathematical theory was proposed, based on topology over the space of infinite sequences (representing the plays in the game), which enables one to reason on sets of observationally equivalent sequences (from the point of view of one player). Significant results for the one-player case, and applications to the diagnosis of discrete event systems have already been obtained; a publication in the 9th International Workshop on Discrete Event Systems is forthcoming.
- Third, optimality criteria on the coalition policies have recently been investigated, based on search techniques of the Artificial Intelligence (AI) field. For 'safety' games, i.e. games where the objective is to avoid 'bad' configurations, whether the observation capabilities of the players are sufficient for a coalition to win the game is in itself a challenging problem. Inspired from control theory, a sufficient condition was considered for the existence of a solution, which is decidable (i.e. which can be checked automatically). The research team studied a way to fix situations where this condition fails: by 'fixing', we mean to add communications between players, thereof augmenting their observation power to make the condition true. Penalising communication (by a cost function) led to solve optimisation problems by AI search techniques: the procedure renders an optimal communication addendum amidst players.
- Fourth, an interdisciplinary collaboration aims at bridging the gap between automated planning techniques and formal methods that are used for software verification. Automated planning is a field of AI. The basic problem consists in automatically choosing and organising a set of actions to reach a given objective in an optimal manner. Application areas include defence, business, and space. In software verification, the basic problem is also to determine whether a state of the software is reachable (typically an undesirable state where a deadlock or an overflow exists), or more generally whether all executions satisfy a given property. The close proximity between planning and verification is evident but has never been the subject of a solid theoretical study; these similarities are analysed as also the differences with an approach under game theory angle.