RL Compiler

Dynamic neural network optimization for edge devices using reinforcement learning to autonomously discover optimal quantization, pruning, and fusion passes.

Edge AI / Neural Compiler

2025–2026

The challenge

Deploying deep neural networks to resource-constrained microcontrollers and edge hardware typically requires weeks of manual tuning. Existing quantization and pruning passes are applied in a fixed order regardless of the underlying model topology, often causing steep accuracy drops or suboptimal latency gains.

The approach

We formulated compiler pass selection as a Markov Decision Process (MDP). A reinforcement learning agent explores pass permutations—interleaving 8-bit/4-bit quantization, structured pruning, and layer fusion—evaluating against a composite reward function balancing top-1 accuracy retention and edge execution latency.

What we made

RL Optimization Pass Agent
Automated IR Transformation Pipeline
Accuracy vs. Latency Reward Referee
Hardware Constraint Benchmarker
Comparative Profiling Dashboard

Built with

Python / PyTorch / Reinforcement Learning (PPO/DQN) / ONNX Runtime / TinyML Toolchain

Credits

Designed and engineered as a core systems and machine learning experiment by the Certified Cool team.