Deep Reinforcement Learning for Real-Time Strategy Games Techniques and Open Challenges

¹ Department of Physics, School of Sciences and Humanities, SR University, Warangal, Telangana, India
² Department of Computer Applications, T. John College, Bangalore, Karnataka, India
³ Department of Computer Science and Engineering, MLR Institute of Technology, Hyderabad, Telangana, India
⁴ Professor, Department of Computer Science and Engineering, CVR College of Engineering, Hyderabad, Telangana, India
⁵ Assistant Professor, Department of Computer Science and Engineering, Bharati Vidyapeeth’s College of Engineering, A-4, Rohtak Road, Paschim Vihar, Delhi, India
⁶ Assistant Professor, Department of CSE, New Prince Shri Bhavani College of Engineering and Technology, Chennai, Tamil Nadu, India

psm45456@gmail.com
felcy_judith@yahoo.com
vtmurthy.v@gmail.com
drddurgabhavani@gmail.com
mohit.t.bvcoe@gmail.com
thangam.v@newprinceshribhavani.com

Abstract

The RTS games are one of the incredibly challenging tasks for the AI because of their large action space, long-term strategic planning, and multi-agent cooperation requirements. Conventional deep reinforcement learning (DRL) methods are effective but may have limitations in terms of scalability, computational grandiosity, generalization capabilities, and interpretable analyses. We present a deep reinforcement learning framework that overcomes these hurdles by boosting multi-agent coordination, sample efficiency, and employing explainable AI (XAI) techniques to improve the model interpretability in rigorous decision-making. In contrast to these existing methods, which rely on large amounts of computation and are severely limited in long-term strategic adaptation, our design features hierarchical learning, curriculum to shape rewards across adjudicated proxy games, and Bayesian uncertainty to promote work in action areas consistent with changing dynamics relative to game mechanics; thereby facilitating rapid adaptability to new situations in RTSs. We also propose dynamic action pruning methods to alleviate redundant action space representation, as well as enhancing the advantage of real-time decision-making. We validate our proposed model over diverse RTS environments, and it not only generalizes better but trains faster while having a richer strategic depth than existing state-of-the-art DRL models. This study closes the gap between theoretical advancements and practical RTS applications, introducing an efficient, interpretable and scalable solution for RTS game strategies driven by AI.