Reinforcement Learning for Execution under Dynamic Fees in a Closed-Loop DEX Simulator

Battle Count: 70

Highly relevant to quantitative trading in the DeFi/DEX space. Directly addresses optimal execution under dynamic fees, a growing concern as AMMs adopt trader-facing fee mechanisms. The evaluation discipline (benchmark ladder, frozen reserved blocks, paired bootstrap CIs, transfer-vs-re-solving) is methodologically valuable for any execution research. The finding that model-free RL beats tuned heuristics specifically under dynamic-fee repricing is actionable for DEX execution agents. However, results are model-conditioned and not deployable profit claims, limiting direct trading application.

Moderate-to-high complexity. Requires building a closed-loop market simulator with constant-product pools, dynamic fee rules, noise flow routing, and CEX-AMM arbitrage. The DQN itself is simple (16-64-64-8 MLP), but the evaluation discipline is extensive: multiple benchmark policies, validation-tuned hyperparameters, frozen reserved blocks, paired bootstrap inference, completion audits, priority ordering permutations, and sensitivity layers. The simulator is implemented in Rust with a PyTorch DQN bridge. Training is CPU-only (12,000 episodes) but the full evaluation pipeline with 1,000+ seeds across multiple orderings and modes is computationally non-trivial.

Exceptionally reproducible: code available on GitHub (https://github.com/egpivo/amm-lab), simulator is bit-deterministic per seed (content-hash verified), CPU-only single-threaded training, all hyperparameters and seed protocols fully disclosed, run manifest with SHA-256 hashes of checkpoints and result files, observation schema whitelist test, exact shortfall decomposition identity regression test, no RL framework used (custom Rust simulator + PyTorch DQN), Gymnasium-style interface over JSON lines. Reserved evaluation block (seeds 90,000-90,999) never touched during development.