A Comparison of Case-Based Reasoning and Probabilistic Graphical Models in the Context of Learning from Observation

Abstract

Learning from observation is a technique whereby learning occurs through observation or experience. In this work, we compare two existing techniques of learning from observation: Probabilistic Graphical Models (PGM) and Case-Based Reasoning (CBR) with the goal of identifying a preferred approach for future improvement. We show that the Naive Bayes Classifier is better than a previously used PGM model in learning behavior in a vacuum cleaner domain and introduce a state-based retrieval technique in CBR and show that there is no once-size-fits-all approach to learn state-based behavior. We also compare the two learning techniques in fully and partially observable continuous domains, namely Cartpole V0, obstacle avoidance, and 2D RoboCup. We show that the CBR approach works best in Cartpole V0, the PGM approach works best in obstacle avoidance, and the PGM approach works best in 2D RoboCup. Ultimately, we show that the preferred technique is generally behavior and domain specific. i