Towards Adaptive Role Selection for Behavior-Based Agents Danny Weyns, Kurt Schelfthout, Tom Holvoet, and Olivier Glorieux AgentWise, DistriNet, Department of Computer Science, K.U.Leuven, Celestijnenlaan 200A, B-3001 Leuven, Belgium {Danny.Weyns, Kurt.Schelfthout, Tom.Holvoet}@cs.kuleuven.ac.be Abstract. This paper presents a model for adaptive agents. The model describes the behavior of an agent as a graph of roles, in short a behavior graph. Links between roles provide conditions that determine whether the agent can switch roles. The behavior graph is assigned at design time, however adaptive role selection takes place at runtime. Adaptivity is achieved through factors in the links of the behavior graph. A factor models a property of the agent or its perceived environment. When an agent can switch roles via different links, the factors determine the role the agent will switch to. By analyzing the effects of its performed actions the agent is able to adjust the value of specific factors, adapting the selec- tion of roles in line with the changing circumstances. Models for adaptive agents typically describe how an agent dynamically selects a behavior (or action) based on the calculation of a probability value as a function of the observed state for each individual behavior (or action). In contrast, the model we propose aims to dynamically adapt logical relations between different behaviors (called roles here) in order to dynamically form paths of behaviors (i.e. sequences of roles) that are suitable in the current state. To verify the model we applied it to the Packet-World. In the paper we discuss simulation results that show how the model enables the agents in the Packet-World to adapt their behavior to changes in the environment. 1 Introduction Adaptability is a system’s capacity to take into account unexpected situations. Multi-agent systems in particular are characterized by the property that not everything can be anticipated in advance. In the context of cognitive agent sys- tems the problem of adaptation is tackled by introducing learning techniques. Traditional learning techniques are based on the agents’ symbolic models of the environment. This however, does not fit the approach of behavior-based agents since these agents do not build a symbolic model of their environment [1, 2, 3]. To deal with the problem of adaptation, an agent has to take into account the quality of the effects realized by its past decisions. Several techniques for behavior-based agent architectures have been proposed to realize adaptation, some examples are [4, 5, 6]. In our research group we also developed two archi- tectures for adaptive agents [7, 8]. All these models typically describe how an D. Kudenko et al. (Eds.): Adaptive Agents and MAS II, LNAI 3394, pp. 295–312, 2005. c  Springer-Verlag Berlin Heidelberg 2005