Agent Ordering and Nogood Repairs in Distributed Constraint Solving

Abstract

The distributed constraint satisfaction problem is a general formalization used to represent problems in distributed multi-agent systems. A large body of problems in artificial intelligence and computer science can be easily formulated as distributed constraint satisfaction problems. In this thesis we study agent ordering, effects of no-goods, search efficiency and threshold repairing in distributed constraint satisfaction problems and its variants. A summary of contributions is as follows: 1. We present a new algorithm, Dynamic Agent Ordering. A distinctive feature of this algorithm is that it uses the degree of unsatisfiability as a guiding parameter to dynamically determine agent ordering during the search. We show through an empirical study that our algorithm performs better than the existing approaches. In our approach, the independence of agents is guaranteed and agents without neighbouring relationships can run concurrently and asynchronously. (Part of this work was published in the Australian Al Conference (80)). 2. We extend the Dynamic Agent Ordering algorithm by incorporating a novel technique called nogood repairing. This results in a dramatic reduction in the nogoods being stored, and communication costs. In an empirical study, we11 show that this approach outperforms an equivalent static ordering algorithm and a current state-of-the-art technique in terms of execution time, memory usage and communication cost. (Part of this work was published at FLAIRS Conference (81)). Further, we introduce a new algorithm, Over-constrained Dynamic Agent Ordering, that breaks new ground in handling multiple variables per agent in distributed over-constrained satisfaction problems. The algorithm also uses the degree of unsatisfiability as a measure for relaxing constraints, and hence as a way to guide the search toward the best optimal solution(s). By applying our Threshold Repair method, we can solve a distributed constraint satisfaction problem without knowing whether the problem is under- or over-constrained. In an experimental study, we show that the new algorithm compares favourably to an implementation of asynchronous weak commitment search adapted to handle over-constrained problems. (Part of this work was published at the Canadian AI conference (79)).