Robert Junges

Evaluation of techniques for a learning-driven modeling methodology in multiagent simulation

There have been a number of suggestions for methodologies supporting the development of multiagent simulation models. In this contribution we are introducing a learning-driven methodology that exploits learning techniques for generating suggestions for agent behavior models based on a given environmental model. The output must be human-interpretable. We compare different candidates for learning techniques - classifier systems, neural networks and reinforcement learning - concerning their appropriateness for such a modeling methodology.

Evolution for modeling: a genetic programming framework for sesam

Developing a valid agent-based simulation model is not always straight forward, but involves a lot of prototyping, testing and analyzing until the right low-level behavior is fully specified and calibrated. Our aim is to replace the try and error search of a modeler by adaptive agents which learn a behavior that then can serve as a source of inspiration for the modeler. In this contribution, we suggest to use genetic programming as the learning mechanism. For this aim we developed a genetic programming framework integrated into the visual agent-based modeling and simulation tool SeSAm, providing similar easy-to-use functionality.

Generating inspiration for agent design by reinforcement learning

One major challenge in developing multiagent systems is to find the appropriate agent design that is able to generate the intended overall dynamics, but does not contain unnecessary features. In this article we suggest to use agent learning for supporting the development of an agent model during an analysis phase in agent-based software engineering. Hereby, the designer defines the environmental model and the agent interfaces. A reward function captures a description of the overall agent performance with respect to the intended outcome of the agent behavior. Based on this setup, reinforcement learning techniques can be used for learning rules that are optimally governing the agent behavior. However, for really being useful for analysis, the human developer must be able to review and fully understand the learnt behavior program. We propose to use additional learning mechanisms for a post-processing step supporting the usage of the learnt model.

Learning Tools for Agent-Based Modeling and Simulation

In this project report, we describe ongoing research on supporting the development of agent-based simulation models. The vision is that the agents themselves should learn their (individual) behavior model, instead of letting a human modeler test which of the many possible agent-level behaviors leads to the correct macro-level observations. To that aim, we integrate a suite of agent learning tools into SeSAm, a fully visual platform for agent-based simulation models. This integration is the focus of this contribution.