Using reinforcement learning for similarity assessment in case-based systems

This research adds reinforcement learning to the similarity assessment component of case-based reasoning systems (CBR). The argument is that reinforcement learning allows case-based reasoning to be more easily used in knowledge-poor domains.; The system uses a case base composed of cases that were previously encountered. Each case suggests a solution to be tried in a particular context. In an application, the system is presented with the context for a new problem situation. The system finds a similar context within its case base and returns the suggested solution. The user then provides the system with a reinforcement signal which indicates how well the solution performs in the new context. Given this feedback, the system modifies how it determines if a case is appropriate for a given context. Given enough contexts and reinforcement signals, the system learns which cases are appropriate for which situations.; A system is implemented that, given current weather conditions, predicts future conditions. This system demonstrates that the method is successful in learning distance metrics. Experiments have also been conducted to compare this approach to other techniques that have been applied to the same applications. One of these alternative approaches is to use the current context as the input to a feed-forward neural net. The network is trained with back-propagation to output the best solution for a particular context. Results show that the system performs significantly better than this approach.; An adaptation mechanism for CBR systems is also presented. This mechanism is designed to work in knowledge-deficient domains. Nearest-neighbor techniques are used to calculate output values using the output of the nearest neighbor plus a vector calculated by using a linear combination of the difference of the current case and its nearest neighbor.