|
Organization responsible: University of Amsterdam People involved: Project description: The project aims to research autonomic capabilities for synthetic multi-agent systems: the capability of synthetic systems to manage themselves, heal themselves and regulate themselves with minimal human intervention.
The subject of our research are embodied agents; agents that are part of a real world application. In this real world the agents have to make correct decision about their actions. When multiple agents are active, the task to find an optimal plan becomes even more challenging (depending on the number of agents, their correlations and the planning horizon). In the real world we cannot assume that the state can be unambiguous determined from the observations, nor that the effect of the actions is fully deterministic; the agents live in a stochastic world. By exploring their action space in an intelligent way, multi-agent systems can learn the future reward they can expect. In order to function optimally with regards to some objective function, the agents must have a good level of situation awareness and adequately perform coordination and planning, the complexity of which increases with the number of agents, their correlations and the planning horizon.
In real life, this task is particularly difficult. Much research in multiagent systems (MAS) makes the simplifying assumption that the world can be observed and that effects of actions are deterministic. In real the real world, however, the environment in which a team of agents is situated is usually stochastic and only partially observable, for example because of sensor noise or perceptual aliasing (in the real world we cannot assume that the state can be unambiguous determined from the observations). In this case the decision making task becomes much harder. In this setting each agent will have to form a belief about the state of the world as well as over the beliefs of other agents, in order to predict their actions.
The research aims at study and design of (approximate) models and algorithms for to enable robust reasoning in systems with multiple independent agents. Multi-agent systems are often been advocated as methodology because multiple agents offer redundancy. In this research we investigate formal models for representing decision problems faced by teams of agents in an uncertain, stochastic world. In particular, we focus on the decentralized partially observable Markov decision process (Dec-POMDP), a framework for fully cooperative planning by a team of agents in such world. A Dec-POMDP is fully cooperative per definition, as it models a single reward function that applies for all agents, i.e., the reward is shared by all agents. As a result, the agents try to optimize the same thing and thus are fully cooperative. Other models in, which there are multiple individual reward functions, can model self-interested agents. An example is the partially observable stochastic game (POSG). Although such models will initially not be our main focus, they are very relevant and will receive more attention when opportunities arise or when this is beneficial from the project perspective. Publications:
|
