Combining search strategies for distributed constraint satisfaction.
Magaji, Amina Sambo-Muhammad
MetadataShow full item record
Many real-life problems such as distributed meeting scheduling, mobile frequency allocation and resource allocation can be solved using multi-agent paradigms. Distributed constraint satisfaction problems (DisCSPs) is a framework for describing such problems in terms of related subproblems, called a complex local problem (CLP), which are dispersed over a number of locations, each with its own constraints on the values their variables can take. An agent knows the variables in its CLP plus the variables (and their current value) which are directly related to one of its own variables and the constraints relating them. It knows little about the rest of the problem. Thus, each CLP is solved by an agent which cooperates with other agents to solve the overall problem. Algorithms for solving DisCSPs can be classified as either systematic or local search with the former being complete and the latter incomplete. The algorithms generally assume that each agent has only one variable as they can solve DisCSP with CLPs using “virtual” agents. However, in large DisCSPs where it is appropriate to trade completeness off against timeliness, systematic search algorithms can be expensive when compared to local search algorithms which generally converge quicker to a solution (if a solution is found) when compared to systematic algorithms. A major drawback of local search algorithms is getting stuck at local optima. Significant researches have focused on heuristics which can be used in an attempt to either escape or avoid local optima. This thesis makes significant contributions to local search algorithms for DisCSPs. Firstly, we present a novel combination of heuristics in DynAPP (Dynamic Agent Prioritisation with Penalties), which is a distributed synchronous local search algorithm for solving DisCSPs having one variable per agent. DynAPP combines penalties on values and dynamic agent prioritisation heuristics to escape local optima. Secondly, we develop a divide and conquer approach that handles DisCSP with CLPs by exploiting the structure of the problem. The divide and conquer approach prioritises the finding of variable instantiations which satisfy the constraints between agents which are often more expensive to satisfy when compared to constraints within an agent. The approach also exploits concurrency and combines the following search strategies: (i) both systematic and local searches; (ii) both centralised and distributed searches; and (iii) a modified compilation strategy. We also present an algorithm that implements the divide and conquer approach in Multi-DCA (Divide and Conquer Algorithm for Agents with CLPs). DynAPP and Multi-DCA were evaluated on several benchmark problems and compared to the leading algorithms for DisCSPs and DisCSPs with CLPs respectively. The results show that at the region of difficult problems, combining search heuristics and exploiting problem structure in distributed constraint satisfaction achieve significant benefits (i.e. generally used less computational time and communication costs) over existing competing methods.