Simulation and analysis of adaptive agents: An integrative modelling approach

Agent-based simulation methods are a relatively new way to address complex systems. Usually, the idea is that the agents used are rather simple, and the complexity and adaptivity of such a system are modeled by the interaction between these agents. However, another way to exploit agent-based simulation methods is by use of agents that themselves also have certain forms of learning or adaptation. In order to simulate adaptive agents with abilities matching those of their real-world biological or societal counterparts, a natural approach is to incorporate certain adaptation mechanisms such as classical conditioning into agent models. Existing models for adaptation mechanisms are usually based on quantitative, numerical methods, and in particular, differential equations. Since agent-based simulation is usually based on qualitative, logical languages, these quantitative models are often not directly appropriate as an input in the context of agent-based simulation. To deal with this problem, this paper puts forward an integrative approach to simulate and analyze the dynamics of complex systems, in particular a conditioning process of an adaptive agent, integrating quantitative, numerical and qualitative, logical aspects within one expressive temporal specification language. To obtain a simulation model, an executable sublanguage of this language is used to specify the agent's adaptation mechanism in detail. For analysis and validation, in the proposed approach both properties characterising the externally observable adaptive behavior and properties characterizing the dynamics of internal intermediate states have been identified, formally specified and automatically checked on the generated simulation traces. As part of the latter, an approach to (formally) specify and check representational relations for intermediate, internal agent states is put forward. This enables verification of whether the representational content of an intermediate state a modeller has in mind indeed is in accordance with the agent model's internal dynamics. For a biological agent with known neural mechanisms, such as Aplysia, the modeling approach incorporates high-level modeling of neural states occurring as intermediate states and relates them to their representational content specification. This provides the possibility to validate not only the resulting observable behavior of a simulation model against the observable behavior of the agent in the real world, but also the intermediate states of the agent in the model against the intermediate states of the agent in the world