Intrinsically Motivated Hierarchical Skill Learning in Structured Environments

摘要: We present a framework for intrinsically motivated developmental learning of abstract skill hierarchies by reinforcement agents in structured environments. Long-term can drastically improve an agent's efficiency solving ensembles related tasks complex domain. In domains composed many features, understanding the causal relationships between actions and their effects on different features environment greatly facilitate learning. Using Bayesian network structure (learning techniques dynamic programming algorithms), we show that learn incrementally autonomously both hierarchy skills exploit this structure. Furthermore, novel active scheme employs intrinsic motivation to maximize with which is learned. As new acquired using current set skills, more are learned, turn allow agent discover structure, so on. This bootstrapping property makes our approach process results steadily increasing domain knowledge behavioral complexity as continues explore its environment.