Conclusion

Abstract

Although the tradeoff between search and knowledge is a well known phenomenon in AI, this dissertation is perhaps the first to investigate how the tradeoff effects the utility of learning. The primary claim, or the “thesis of the thesis”, is that a learning system must be sensitive to the knowledge/search tradeoff if it is to have a positive influence on performance. Figure 13–1, adapted from Ebeling [22], graphically illustrates the prototypical relationship between knowledge and search speed (e.g., time per node expansion). The curves represent a constant level of performance over a range in knowledge and search speed. The arrow labeled “positive utility” indicates the best case that a learning program can achieve; knowledge is added to the system without a corresponding decrease in search speed, resulting in improved performance. The arrow label “negative utility” represents the converse case where adding knowledge impairs performance, an all too likely possibility. The results of my research indicate that an EBL system can reliably improve performance by searching for effective control knowledge. The choice of target concept and training example, the representation of the resulting control knowledge, and the determination of whether the control knowledge is actually useful all play an important role in EBL.