Assisted Feature Engineering and Feature Learning to Build Knowledge-based Agents for Arcade Games

Human knowledge can greatly increase the performance of autonomous agents. Leveraging this knowledge is sometimes neither straightforward nor easy. In this paper, we present an approach for assisted feature engineering and feature learning to build knowledge-based agents for three arcade games within the Arcade Learning Environment. While existing approaches mostly use model-free approaches we aim at creating a descriptive set of features for world modelling and building agents. To this end, we provide (visual) assistance in identifying and modelling features from RAM, we allow for learning features based on labeled game data, and we allow for creating basic agents using the above features. In our evaluation, we compare different methods to learn features from the RAM. We then compare several agents using different sets of manual and learned features with one another and with the state-of-the-art.