Adaptive Dueling Double Deep Q-networks in Uniswap V3: Replication and Extension with Mamba

Battle Count: 74

Highly relevant to quantitative trading in DeFi/AMM contexts. The paper addresses adaptive liquidity provision strategies, risk management (LVR), and market making optimization. The Mamba-DDQN extension introduces temporal modeling for financial time series. However, the work is primarily a replication study with a preliminary extension, and results are not yet fully validated across all datasets. The reward function design incorporating gas fees, trading fees, and LVR is directly applicable to DeFi trading strategies.

Moderate to high complexity. Requires understanding of: (1) Uniswap V3 concentrated liquidity mechanics and tick-based pricing; (2) Dueling DDQN architecture with experience replay; (3) Mamba State Space Model implementation; (4) Feature engineering from blockchain data (28+ technical indicators); (5) Custom reward function design; (6) Multiple baseline implementations (Uniform τ-reset, EWA, Dynamic Programming). The code snippets suggest Python/PyTorch implementation. Computational requirements are significant (>20 min/epoch on RTX 2070).

The paper explicitly discusses significant reproducibility challenges: (1) Data differences due to time lag and missing hourly volume data from The Graph; (2) RL stochasticity from unspecified seeds, training epochs, and early stopping details; (3) Missing hyperparameters like exact gas fee values; (4) Unclear instructions about hedging variants; (5) Computational resource constraints (RTX 2070, >20 min/epoch); (6) Authors did not respond to email inquiries. The original paper's link was unavailable, requiring alternative data sourcing.