How Much Can Experimental Cost Be Reduced in Active Learning of Agent Strategies?

• Published in: Inductive Logic Programming Vol. 11105; pp. 38 - 53
• Main Authors: Hocquette, Céline, Muggleton, Stephen
• Format: Book Chapter
• Language: English
• Published: Switzerland Springer International Publishing AG 01.01.2018; Springer International Publishing
• Series: Lecture Notes in Computer Science
• Subjects: Active learning; Agent-based modelling; Bayesian meta-interpretive learning
• ISBN: 9783319999593; 3319999591
• ISSN: 0302-9743; 1611-3349
• DOI: 10.1007/978-3-319-99960-9_3

In science, experiments are empirical observations allowing for the arbitration of competing hypotheses and knowledge acquisition. For a scientist that aims at learning an agent strategy, performing experiments involves costs. To that extent, the efficiency of a learning process relies on the number of experiments performed. We study in this article how the cost of experimentation can be reduced with active learning to learn efficient agent strategies. We consider an extension of the meta-interpretive learning framework that allocates a Bayesian posterior distribution over the hypothesis space. At each iteration, the learner queries the label of the instance with maximum entropy. This produces the maximal discriminative over the remaining competing hypotheses, and thus achieves the highest shrinkage of the version space. We study the theoretical framework and evaluate the gain on the cost of experimentation for the task of learning regular grammars and agent strategies: our results demonstrate the number of experiments to perform to reach an arbitrary accuracy level can at least be halved.