Dynamic Skill Learning: A Support to Agent Evolution

In this paper, we show that every agent can be built from an atomic agent through dynamic skill a quisition, a skill being a coherent set of abilities. At first, we propose a definition of an atomic agentand then we present the skill notion. In our work, since an agent is defined by the set of roles he can play according to the skills he has learned, we can consider that skills are the backbone of agents. We propose that the learning be dynamic and thus agents can effectively evolve during their “life”. That means that the roles he plays can change. This approach promotes evolutivity, reusability and modularity. These concepts have been applied to our own multi-agent system model, called M AGIQUE. It is based on hierarchies of agents (or agents recursively built from agents depending on the vision). This organisation provides an automatic delegation of the exploitation of skills between agents, that contributes to the simplicity and adaptability of MAS building. An API that implements these ideas has been developed and of the corresponding API that. It provides an easy to use framework to build multi-agent applications where agents effectively dynamically evolve.