Mastersthesis,
Decision Making in a Multi-Agent Planning versus Learning
Graz University of Technology, Graz, Austria, (January 2008)
Abstract
Within the research field of autonomous agents, decision making is a central topic. In this thesis two different approaches - one by planning and an alternative one via learning are described in their application. While the current implementation of the planning system is based on a Prolog backend, the learning approach is implemented with a reinforcement Q-learning algorithm. The environment of the application is the system of KickOffTUG, the RoboCup 2D Soccer Simulation League team of Graz, University of Technology. Within this work both approaches are compared and evaluated to discover their strengths and weaknesses. It could be shown that the main advantage of the planning system is that the plans are built and reviewed by humans. Unfortunately the dependency on the human is the biggest problem of this system as well. On the other hand the learning system is more adaptable, but its main disadvantage lies in the specification of the target function and the creation of the training examples. This work shows that both algorithms, planning and learning, can be applied to the decision making process, as they are equal in their performance.
