arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 1502.03248cs.AI
收藏

非政策奖励塑造与集成

Off-Policy Reward Shaping with Ensembles

Vrije Universiteit Brussel(布鲁塞尔自由大学)

Anna Harutyunyan, Tim Brys, Peter Vrancx, Ann Nowe

摘要:

基于潜力的奖励塑造(PBRS)是一种有效且流行的技术,通过利用领域知识来加速强化学习。虽然PBRS被证明始终保持最优策略,但其对学习速度的影响取决于其潜在函数的质量,而潜在函数的质量又取决于底层启发式和规模。要知道哪种启发式算法将被证明是有效的,需要事先测试选项,并且确定适当的尺度需要调整,这两者都引入了额外的样本复杂性。我们制定了一个PBRS框架,降低学习速度,但不会产生额外的样本复杂性。为此,我们建议同时学习一套政策,形成w.r.t.许多化学品和一系列规模。然后通过投票获得目标策略。集成需要能够有效和可靠地学习离线策略:最近的部落架构满足的要求,我们将其作为我们的基础。我们的经验表明,(1)我们的合奏政策优于基本政策,其单一的启发式组件,和(2)合奏在一般范围内的规模执行至少以及一个最佳调整组件。

英文摘要:

Potential-based reward shaping (PBRS) is an effective and popular technique to speed up reinforcement learning by leveraging domain knowledge. While PBRS is proven to always preserve optimal policies, its effect on learning speed is determined by the quality of its potential function, which, in turn, depends on both the underlying heuristic and the scale. Knowing which heuristic will prove effective requires testing the options beforehand, and determining the appropriate scale requires tuning, both of which introduce additional sample complexity. We formulate a PBRS framework that reduces learning speed, but does not incur extra sample complexity. For this, we propose to simultaneously learn an ensemble of policies, shaped w.r.t. many heuristics and on a range of scales. The target policy is then obtained by voting. The ensemble needs to be able to efficiently and reliably learn off-policy: requirements fulfilled by the recent Horde architecture, which we take as our basis. We demonstrate empirically that (1) our ensemble policy outperforms both the base policy, and its single-heuristic components, and (2) an ensemble over a general range of scales performs at least as well as one with optimally tuned components.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑