A Performance Measure for Intelligentmachines Based on Complexity And

This paper introduces a formalism which combines reliability and complexity in a measure of performance for Intelligent Machines. For a given desired accuracy, diierent algorithms may be available which are reliable enough. Reliability is deened as the probability that an algorithm meets, at run-time, the accuracy speciications which guided its design. However it is important to have a means of choosing the algorithm of least cost among the reliable ones. The cost measures resources usage, such as CPU time, memory space or number of processors. The Theory of Information-Based Complexity provides a formalism to deal with diierent sources of information, thus with distinct algorithms at all levels of the machine. Two case studies on motion control and image processing illustrate the method.