The Case for Case-Based Transfer Learning

54 AI MAGAZINE Observations of human reasoning motivate AI research on transfer learning (TL) and case-based reasoning (CBR). Our ability to transfer knowledge and expertise from understood domains to novel ones has been thoroughly documented in psychology and education (for example, Thorndike and Woodworth 1901; Perkins and Salomon 1994; Bransford, Brown, and Cocking 2000), among other disciplines. Transfer learning uses knowledge learned from solving tasks from a source domain to enhance an agent’s ability to learn to solve tasks from a target domain. The differences between the source and target problems characterize the transfer distance. Casebased reasoning transfers problem-solving knowledge from specific examples or episodes, called cases, to new problems. While researchers typically work within each of these fields independently, the purpose of this article is to summarize how case-based reasoning can be applied to transfer learning.1 Our analysis reveals three approaches for applying CBR to transfer learning: (1) CBR as a transfer learning method, (2) CBR for problem solving, and (3) CBR to transfer knowledge between the domains. These correspond to using CBR for solving entire transfer tasks or acting as a component within a transfer learning system. The transfer distance provides a new metric for CBR researchers to assess the robustness of their systems. Furthermore, each CBR approach has implications for the importance and interpretations of different transfer learning metrics. We begin with an overview of transfer learning, case-based reasoning, and the three approaches for applying CBR to transfer learning. We provide examples of each from our research on physics problem solving and controlling a player in a football simulation and also describe related work. We close with a discussion concerning the applicability of each approach over different transfer distances, how transfer learning assists CBR research, and some directions for future research.