Andraž Bežek (2007) Automatic Modeling of Multiagent. PhD thesis.
One of the most difficult unsolved tasks in the field of multi-agent modeling is to discover common agent strategy by knowing only low-level agent behavior and basic domain knowledge. This task is difficult for the following two reasons. First, agents are autonomous entities that try to accomplish in advance given strategy, while continuously interacting with team-agents, adversary-agents, and the environment. By giving only low-level descriptions of single agent behaviors, high-level strategy is therefore difficult to discover. Second, in addition to low-level agent behavior, the system utilizes only basic domain knowledge. Such knowledge is often available and the task is to discover high-level knowledge, usually available only to experts. This dissertation presents a novel algorithm for strategic modeling MASDA (Multi-Agent Strategy Discovering Algorithm). By tracking low-level behavior of a group of agents and using only basic domain knowledge, it tries to discover common agent strategy. MASDA incorporates an abstraction process, which enables MASDA to separate models that are due to following of agent strategy, from models that are due to agent reactions to local environment changes. The created model describes collaboration of agents in a graphical and symbolical manner. Graphic descriptions visually present multi-agent activity, while symbolic descriptions present important characteristics of multi-agent interaction in a human-comprehendible way. By that human analysis, interpretation and automatic classification is possible. Efficiency of the algorithm was shown on two multi-agent domains. In the domain of robotic soccer – RoboCup, the human expert confirmed that the created model semantically describes part of real-life soccer strategy. Measurements of classification accuracy, recall, and precision confirm suitability of MASDA models for computer classification. In the domain of ball keep away problem – 3vs2 Keepaway, the declarative efficiency of the model was confirmed by comparing it with source code for three known strategies and the strategy imitation performance was shown by measuring time efficiency and by human and computer comparison of played actions.
Actions (login required)