|
SNN-3: Modelling and Learning of Graphical Multi-Agent Systems |
|
Organization responsible: SNN Adaptive Intelligence People involved: Project description: This project aims to investigate modelling and learning in graphical models representing multi-agent systems, which should lead to new means to model collaborative reasoning in dynamic large-scale agent organizations.
Single agent decision making under uncertainty has been successfully modelled by influence diagrams and Markov decision processes. Recently, several generalizations of these models have been proposed to describe collaborative decision making in multi-agent systems (e.g. the multi-agent influence diagrams and the multi-agent Markov decision processes). The common denominator in these approaches is that the system is represented as a sparse graph with limited local interactions. These representations can be viewed as generalizations of the Bayesian network representation of probabilistic models or more precisely, as generalizations of influence diagrams – also known as decision networks. Basically, a multi-agent system is described by a set of partly overlapping influence diagrams, where each agent has its diagram with access to its own stochastic nodes, action nodes and utility nodes. In this framework, the system is fully defined and can be used to compute, for instance, the optimal action of an agent (which, in turn, is in general a computationally non-trivial task. This however is addressed in another subproject).
An important question is, however, how such models are constructed in the first place. This is an issue in collaborative multi-agent systems but it is also still an important issue in real world applications of `standard’ Bayesian networks. The main research question is how modelling and learning in systems that support or make decisions under uncertainty can be efficiently achieved. Publications:
|
