Learn to human-level control in dynamic environment using incremental batch interrupting temporal abstraction.

The temporal world is characterized by dynamic and variance. A lot of machine learning algorithms are difficult to be applied to practical control applications directly, while hierarchical reinforcement learning can be used to deal with them. Meanwhile, it is a commonplace to have some partial solutions available, called options, which are learned from knowledge or predefined by the system, to solve sub-tasks of the problem. The option can be reused for policy determination in control. Many traditional semi-Markov decision process methods take advantage of it. But most of them treat the option as a primitive object. However, due to the uncertainty and variability of the environment, they are unable to deal with real world control problems effectively. Based on the idea of interrupting option under the prerequisite for dynamic environment, a Q-learning control method which uses temporal abstraction, named as I-QOption, is introduced. The I-QOption approach combines the idea of interruption with the characteristics of dynamic environment so as to be able to learn and improve control policy in dynamic environment. The Q-learning framework helps to learn from interaction with raw data and achieving human-level control. The I-QOption algorithm is applied to grid world, a benchmark dynamic environment evaluation testing. The experiment results show that the proposed algorithm can learn and improve policy effectively in dynamic environment.