Search for a command to run...
Expert systems divide neatly into two categories: those in which ( 1) the expert decisions result in changes to some external environment (control systems), and (2) the expert decisions merely seek to describe the environment (classification systems). Both the explanation of computer-based reasoning and the "bottleneck" (Feigenbaum, 1979) of knowledge acquisition are major issues in expert systems research. We have contributed to these areas of research in two ways. Firstly, we have implemented an expert system shell, the Mugol environment, which facilitates knowledge acquisition by inductive inference and provides automatic explanation of run-time reasoning on demand. RuleMaster, a commercial version of this environment, has been used to advantage industrially in the construction and testing of two large classification systems. Secondly, we have investigated a new technique called sequence induction which can be used in the construction of control systems. Sequence induction is based on theoretical work in grammatical learning. We have improved existing grammatical learning algorithms as well as suggesting and theoretically characterising new ones. These algorithms have been successfully applied to the acquisition of knowledge for a diverse set of control systems, including inductive construction of robot plans and chess end-game strategies.