5. Conclusions

negative effect is naturally more significant in the more complex domain. The graph for the simple domain crosses the 0 line earlier than the complex domain. That means that learning starts to be useful with weight greater than 0.6 for the simple domain and 0.7 for the complex domain. As we relax the optimality requirement more significantly (with a W = 0.8), macro usage in the more complex domain becomes more advantageous. The purpose of the research described in this paper is to identify the parameters that effects deductive learning and to perform experiments systematically in order to understand the nature of those effects. The goal of this paper is to demonstrate the methodology of performing parametric experimental study of deductive learning. The example here include the study of two parameters: the point on the satisficing-optimizing scale that is used during the search carried out during problem solving time and during learning time. We showed that A*, which looks for optimal solutions, cannot benefit from macro learning but as the strategy comes closer to best-first (satisficing search), the utility of macros increases. We also demonstrated that deductive learners that learn off-line by solving training problems are sensitive to the type of search used during the learning. We showed that in general optimizing search is best for learning. It generates macros that increase the quality solutions regardless of the search method used during problem solving. It also improves the efficiency for problem solvers that require a high level of optimality. The only drawback in using optimizing search is the increase in learning resources spent. We are aware of the fact that the results described here are not very surprising. The goal of the parametric study is not necessarily to find exciting results, but to obtain results, sometimes even previously known, in a controlled experimental environment. The work described here is only part of our research plan. We are currently in the process of extensive experimentation with all the parameters described here and also with others. We also intend to test the validity of the conclusions reached during the study by repeating some of the tests in several of the commonly known search problems. We hope that such systematic experimentation will help the research community to better understand the process of deductive learning and will serve as a demonstration of the experimental methodology that should be used in machine learning research. Filtering: …