Using chains of bottleneck transitions to decompose and solve reinforcement learning tasks with hidden states
Future Generation Computer Systems(2022)
摘要
Reinforcement learning is known to underperform in large and ambiguous problem domains under partial observability. In such cases, a proper decomposition of the task can improve and accelerate the learning process. Even ambiguous and complex problems that are not solvable by conventional methods turn out to be easier to handle by using a convenient problem decomposition, followed by the incorporation of machine learning methods for the sub-problems. Like in most real-life problems, the decomposition of a task usually stems from the sequence of sub-tasks that must be achieved in order to get the main task done. In this study, assuming that unambiguous states are provided in advance, a decomposition of the problem is constructed by the agent based on a set of chains of bottleneck transitions, which are sequences of unambiguous and critical transitions leading to the goal state. At the higher level, an agent trains its sub-agents to extract sub-policies corresponding to the sub-tasks, namely two successive transitions in any chain, and learns the value of each sub-policy at the abstract level. Experimental study demonstrates that an early decomposition based on useful bottleneck transitions eliminates the necessity for excessive memory and improves the learning performance of the agent. It is also shown that knowing the correct order of bottleneck transitions in the decomposition results in faster construction of the solution.
更多查看译文
关键词
Reinforcement learning,Task decomposition,Chains of bottleneck transitions
AI 理解论文
溯源树
样例
![](https://originalfileserver.aminer.cn/sys/aminer/pubs/mrt_preview.jpeg)
生成溯源树,研究论文发展脉络
Chat Paper
正在生成论文摘要