Development and Implementation of a Case-Based Reasoning Approach to Speed-Up Deep Reinforcement Learning through Case-Injection for AI Gameplay

Game environments offer properties that are useful for researching challenges in artificial intelligence (AI). Gaming enables testing, evaluation, and preparation of new methods for real-world scenarios. Reinforcement learning (RL) has undergone enormous further development in the recent years. The usage of artificial neural networks makes it possible to use reinforcement learning algorithms in complex environments. To learn feasible solutions, RL agents have to interact with the environment and learn based on their experience. Many scenarios require long training times and a vast amount of training data. Reusing previously experience knowledge can be the key to shortened training cycles and improved performance. Case-based reasoning (CBR) is another methodology of artificial intelligence using experiences from previous situations for solving new situations by adapting known solutions. Therefore, CBR appears to be particularly suitable for knowledge transfer in the area of reinforcement learning and is applied to improve the learning process of RL agents within video games. First, this work develops a theoretical approach in order to show in a second step the practical feasibility with the help of a prototypical implementation. The evaluation of the proposed method confirms reduced training time and improved performance.