DL-based prediction of optimal actions of human experts

Jung H. Lee, Ryan S Butner, Elise Saxon, Nathan Oken Hodas

Keywords: deep learning, expert system, sequential learning

Abstract: Expert systems have been developed to emulate human experts’ decision-making. Once developed properly, expert systems can assist or substitute human experts, but they require overly expensive knowledge engineering/acquisition. Notably, deep learning (DL) can train highly efficient computer vision systems only from examples instead of relying on carefully selected feature sets by human experts. Thus, we hypothesize that DL can be used to build expert systems that can learn human experts’ decision-making from examples only without relying on overly expensive knowledge engineering. To address this hypothesis, we train DL agents to predict optimal strategies (actions or action sequences) for the popular game `Angry Birds’, which requires complex problem-solving skills. In our experiments, after being trained with screenshots of different levels and pertinent 3-star guides, DL agents can predict strategies for unseen levels. This raises the possibility of building DL-based expert systems that do not require overly expensive knowledge engineering.