Minutes from “Multi-agent Oriented Constraint Satisfaction”

The paper begins with a definition of a constraint satisfaction problem. Lin noted that we were all quite familiar with this definition, but discussing it helped to use up time in his presentation. The paper describes two general methods for solving a CSP: generate and test (GT), and backtracking search (BT). GT generates possible combinations of variables and checks whether it has found a solution. BT systematically searches for solutions. There are several improvements that have been made to increase the performance of BT. Thrashing can be mitigated by application of consistency techniques. Redundant work can be avoided by using a dependency directed backtracking scheme. Of course, general CSP problems are NP-complete, so even with these enhancements BT is still unable to solve sufficiently large problems. GT can be improved by making use of heuristics and stochastic algorithms. Local search methods typically use iterative repair or hill climbing techniques. To escape from local optima, these techniques will generally make use of a random walk or random restart.