Learning probabilistic relational dynamics for multiple tasks

Abstract

While large data sets have enabled machine learning algorithms to act intelligently in complex domains, standard machine learning algorithms perform poorly in situations in which little data exists for the desired target task. Transfer learning attempts to extract trends from the data of similar source tasks to enhance learning in the target task. We apply transfer learning to probabilistic rule learning to learn the dynamics of a target world. We utilize a hierarchical Bayesian framework and specify a generative model which dictates the probabilities of task data, task rulesets and a common global ruleset. Through a greedy coordinated-ascent algorithm, the source tasks contribute towards building the global ruleset which can then be used as a prior to supplement the data from the target ruleset. Simulated experimental results in a variety of blocks-world domains suggest that employing transfer learning can provide significant accuracy gains over traditional single task rule learning algorithms.