A High-Throughput Compute-Efficient POMDP Hide-And-Seek-Engine (HASE) for Multi-Agent Operations

Reinforcement learning, especially in decentralized partially observable environments (Dec-POMDPs), suffers from high sample complexity and computational cost. To address this, researchers from the University of Tulsa built HASE (Hide-And-Seek-Engine), a compute-efficient engine natively written in C++. It uses Data-Oriented Design principles, explicit 64-byte cache-line alignment to eliminate false sharing, and a zero-copy PyTorch memory bridge via pinned memory and DMA. This design allows HASE to sustain 33,000,000 steps per second (SPS) in a single-agent, 1024-environment setup on an AMD Ryzen 9950X (16 cores). Even with ten agents, throughput only drops to 7M SPS. Compared to a single-threaded vectorized NumPy baseline, this represents a roughly 3,500x speedup.

The engine's performance enables rapid training of cooperative multi-agent policies using standard algorithms like PPO, DQN, and SAC—all completed in minutes. This is a significant leap for simulation-heavy domains such as human-AI joint operations, robotics swarm coordination, and autonomous driving. By optimizing the decision-level simulation layer, HASE makes it feasible to iterate quickly on multi-agent RL experiments without requiring expensive GPU clusters. The paper includes 21 pages, 10 figures, and 5 tables, with code expected to be open-sourced.