Autonomous dynamic reconfiguration in collaborative problem solving

This thesis is a comprehensive study of collaborative problem solving, its advantages and pitfalls, but mainly its improvement by autonomous dynamic reconfiguration. Autonomous dynamic reconfiguration deals with two common problems in collaborative problem solving – high communication effort and poor problem solving results because of restricted overview. The main idea of autonomous dynamic reconfiguration is to autonomously and dynamically adapt the configuration of a collaborative problem solving effort, i.e. the distribution of knowledge, goals and skills, to the problem itself. This is done by two individual local operations – agent melting and agent splitting. Agent melting means to unify the knowledge, goals and skills of two or more agents in a single agent, while agent splitting denotes a process in which a single agent is split, possibly resulting in an additional new agent. In this work we lay a sound theoretical foundation for collaborative problem solving itself and for its improvement by autonomous dynamic reconfiguration. The problem of finding a good configuration for a given problem has been proven to feature a lattice structure and to be complex to solve. Based on a characterization of agent melting and agent splitting as equivalence relation manipulators, these two operators show verifiable properties regarding structure retainment, impact, sufficiency and concurrency. On the practical side, collaborative problem solving finds a very flexible and effective instantiation in this work. All presented concepts for collaborative problem solving, such as the multi-phase agreement finding protocol for external problem solving, the composable belief-desire-intention agent achitecture and the distribution-aware constraint specification architecture for internal problem solving, are designed to support a changing configuration. This set of concepts is integrated by the AuReCon controller that decides by observation and self-observation how to locally adjust the configuration. The proposed theory and practice are validated using a case study in medical appointment scheduling. The autonomous dynamic reconfiguration approach shows convincing results in improving the problem solving quality by up to 20 percent and by more than halving the communication effort. Using autonomous dynamic reconfiguration, the collaborative problem solving techniques can match the quality of a central approach and are more efficient.