Game AI & Applied Machine Learning
Specializing in Deep Reinforcement Learning (PPO/SAC), physics-informed character locomotion, and zero-allocation in-engine neural inference (ONNX, Unreal NNE, LibTorch). Bridging policy training in Python with native C++ game loops.
Formulating continuous action spaces using Proximal Policy Optimization (PPO) and Soft Actor-Critic (SAC). Designing dense/sparse reward functions with curriculum training schedules for non-trivial game tasks.
Integrating trained neural networks into high-frequency engine ticks. Profiling ONNX Runtime and Unreal NNE (Neural Network Engine) with custom C++ memory arenas to eliminate heap allocations per frame.
Leveraging Generative Adversarial Imitation Learning (GAIL) and Behavioral Cloning from human gameplay telemetry to create natural, non-cheating opponent personas and believable boss encounters.
Vectorized continuous aerial combat simulator. Agents learn energy retention, missile evading maneuvers, and cooperative flanking purely through symmetric self-play with an ELO-ranked matchmaking pool.
An open C++20 dynamic module for Unreal Engine 5. Bypasses blueprint overhead by evaluating batch agent sensory tensors in SIMD-aligned arrays, maintaining sub-0.2ms latency across 500 simultaneous actors.
Physics-driven active ragdoll learning terrain traversal from raw mocap clips. Neural policy outputs joint target PD-controller torques, recovering organically from unexpected kinematic impacts.
Pure machine learning often fails in production games because designers lose creative control. My engineering focuses on Hybrid AI architectures: using DRL for complex physical actuation, micro-maneuvers, and high-dimensional perception, while orchestrating high-level tactical intent through transparent hierarchical state machines and designer-tunable influence bounds.