An In-Depth Exploration of Deep Reinforcement Learning Techniques for Game Playing

Abstract

This paper presents a brief survey of the approaches under the DRL umbrella for game playing and a discussion of the problems and the advances that have been witnessed in this promising research line. It is well understandable that the inputs are highly dimensional and the game environment is dynamic and hence the model described above incorporated both CNNs and RNNs. Other components of the methodology were the HRL that started by deconstructing a task into sub-tasks and then the sub-tasks were directed by a master policy. The DRL agents were tested on different platform environments: Specific and steady-build games and platforms and then to evaluate its performance, learning effect, and or its ability to generalize. It was found that the proposed methods were reasonably close to the other effective existing methods; in addition, the time required to train the model was less and the percentage of positive results in the games were also good showing an acceptable generality. The study has also has also extended the prior work on transfer learning as well as the efficient rewarding exploration approach to counter negative impacts of limited rewards and computational time. The data generated by this work contributes to the Doctoral study assessment and is relevant to various DRL themes and can be applied for Robotics and AS fields based on AI. In conclusion, the authors repeat their imperative for future work on expanding DRL approaches that are effective on a large scale, the limitations of application to multi-agent systems, and the transferability of learnt behaviour to improve the AI elements of game playing that is not limited to.