2026/01/24 by Kara Rawson · 1 voice
Computer Science · Decision Sciences · #Advanced Bandit Algorithms Research #Artificial Intelligence in Games #Component (thermodynamics) #Component-based software engineering #Game engine #Markov decision process #Operator (biology) #Reinforcement Learning in Robotics #Reinforcement learning #Software #Transformer
paper · doi:10.5281/zenodo.18359715
openalex publication_date 2026/01/24 · openalex created_date 2026/01/25 · openalex updated_date 2026/07/01
Abstract Short summaryA practical, mathematically grounded framework for implementing reinforcement learning natively inside Entity Component System game engines. The manuscript formalizes MDP and POMDP mappings to ECS components, derives and unifies algorithmic formulations for PPO SAC DQN and CQL, and describes GPU accelerated training via Vulkan compute shaders. It covers memory architectures LSTM GRU Transformer for partial observability, intrinsic motivation modules ICM and RND, Decision Transformer sequence modeling, and production concerns including determinism monitoring and runtime safeguards. The work emphasizes reproducible native C++20 implementations, operator level GPU kernels, and an operational playbook for live training. Why this dataset or software is usefulThis paper bridges the gap between RL research and production game engines by providing implementable pseudocode operator specifications and ECS component schemas that reduce integration friction for engine developers and researchers building real time adaptive agents. Keywordsreinforcement learning; game AI; ECS; Proximal Policy Optimization; Soft Actor Critic; Deep Q Network; Conservative Q Learning; Decision Transformer; intrinsic curiosity; Vulkan; GPU compute; C++; determinism; TBPTT; replay buffer Subjectsartificial intelligence; machine learning; game development; high performance computing; software engineering