Reinforcement learning (RL) is designed to learn optimal control policies from unsupervised interactions with the environment. Many successful RL algorithms have been developed, however, none of them can efficiently tackle problems with high-dimensional state spaces due to the "curse of dimensionality," and so their applicability to real-world scenarios is limited. Here we propose a Sample Aware Feature Selection algorithm embedded in NEAT, or SAFS-NEAT, to help address this challenge. This algorithm builds upon the powerful evolutionary policy search algorithm NEAT, by exploiting data samples collected during the learning process. This data permits feature selection techniques from the supervised learning domain to be used to help RL scale to problems with high-dimensional state spaces. We show that by exploiting previously observed samples, on-line feature selection can enable NEAT to learn near optimal policies for such problems, and also outperform an existing feature selection algorithm which does not explicitly make use of this available data.

Autonomous agents are emerging in diverse areas and many rely on reinforcement learning (RL) to learn optimal control policies by acting in the environment. This form of learning generates large amounts of transition dynamics data, which can be mined to improve the agent's understanding of the environment. There could be many uses for this data, here we focus on mining it to identify a relevant feature subspace. This is vital since RL performs poorly in high-dimensional spaces, such as those that autonomous agents would commonly face in real-world problems. This paper demonstrates the necessity and feasibility of integrating data mining into the learning process while an agent is learning, enabling it to learn to act by both acting and understanding. Doing so requires overcoming challenges regarding data quantity and quality, and difficulty measuring feature relevance with respect to the control policy. We propose the progressive mining framework to address these challenges by relying on cyclic interaction between data mining and RL. We show that a feature selection algorithm developed under this framework, PROFESS, can improve RL scalability better than a competing approach.