Constraint Optimization and Abstraction for Embedded Intelligent Systems

Many tasks in artificial intelligence, such as diagnosis, planning, and reconfiguration, can be framed as constraint optimization problems. However, running constraint optimization within embedded systems requires methods to curb the resource requirements in terms of memory and run-time. In this paper, we present a novel method to control the memory requirements of message-passing algorithms that decompose the problem into clusters and use dynamic programming to compute approximate solutions. It can be viewed as an extension of the previously proposed mini-bucket scheme, which limits message size simply by omitting constraints from the messages. Our algorithm instead adaptively abstracts constraints, and we argue that this allows for a more finegrained control of resources particularly for constraints of higher arity and variables with large domains that often occur in models of technical systems. Preliminary experiments with a diagnosis model of NASA's EO-1 satellite appear promising.