Daniel Sabin

Contradicting Conventional Wisdom in Constraint Satisfaction

Constraint satisfaction problems have wide application in artificial intelligence. They involve finding values for problem variables where the values must be consistent in that they satisfy restrictions on which combinations of values are allowed. Two standard techniques used in solving such problems are backtrack search and consistency inference. Conventional wisdom in the constraint satisfaction community suggests: 1) using consistency inference as preprocessing before search to prune values from consideration reduces subsequent search effort and 2) using consistency inference during search to prune values from consideration is best done at the limited level embodied in the forward checking algorithm. We present evidence contradicting both pieces of conventional wisdom, and suggesting renewed consideration of an approach which fully maintains arc consistency during backtrack search.

Understanding and improving the MAC algorithm

Constraint satisfaction problems have wide application in artificial intelligence. They involve finding values for problem variables where the values must be consistent in that they satisfy restrictions on which combinations of values are allowed. Recent research on finite domain constraint satisfaction problems suggest that Maintaining Arc Consistency (MAC) is the most efficient general CSP algorithm for solving large and hard problems. In the first part of this paper we explain why maintaining full, as opposed to limited, arc consistency during search can greatly reduce the search effort. Based on this explanation, in the second part of the paper we show how to modify MAC in order to make it even more efficient. Experimental results prove that the gain in efficiency can be quite important.

A Constraint-Based Approach to Diagnosing Software Problems in Computer Networks

Distributed software problems can be particularly mystifying to diagnose, for both system users and system administrators. Modelbased diagnosis methods that have been more commonly applied to physical systems can be brought to bear on such software systems. A prototype system has been developed for diagnosing problems in software that controls computer networks. Our approach divides this software into its natural hierarchy of layers, subdividing each layer into three separately modeled components: the interface to the layer above on the same machine, the protocol to the same layer on a remote machine, and the configuration. For each component knowledge is naturally represented in the form of constraints. User interaction modeling is accomplished through the introduction of constraints representing user assumptions, the finite-state machine specification of a protocol is translated to a standard CSP representation and configuration tasks are modeled as dynamic CSPs. Diagnosis is viewed as a partial constraint satisfaction problem (PCSP). A PCSP algorithm has been adapted for use as a diagnostic engine. This paper presents a case study illustrating the diagnosis of some problems involving the widely used FTP and DNS network software.