Relational transfer in reinforcement learning

摘要: Transfer learning is an inherent aspect of human learning. When humans learn to perform a task, we rarely start from scratch. Instead, recall relevant knowledge previous experiences and apply that help us master the new task more quickly. This principle can be applied machine as well. Machine often addresses single tasks in isolation. Even though multiple related may exist domain, many algorithms for have no way utilize those relationships. Algorithms allow successful transfer one (the source) another target) are necessary steps towards making adaptable thesis investigates methods reinforcement (RL), where agent takes series actions environment. RL requires substantial amounts nearly random exploration, particularly early stages The ability therefore important asset agents. source reduce low initial performance common challenging target tasks. I focus on transferring relational guides action choices. Relational typically uses first-order logic express information about relationships among objects. First-order logic, unlike propositional use variables generalize over classes This greater generalization makes effective transfer. contributes six three categories: advice-based transfer, macro MLN Advice-based source-task provide advice target-task learner, which follow, refine, or ignore according its value. Macro-transfer MLN-transfer experience demonstrate good behavior learner. evaluate these experimentally complex reinforcement-learning domain RoboCup simulated soccer. All my empirical benefits compared non-transfer approaches, either by increasing enabling faster task.