Journal of Advances in Civil and Mechanical Engineering
ISSN: 3067-2457 | DOI: 10.64030/3067-2457
Franco Farrugia and Cedric Deguara
Concentrated automated market makers (cAMMs), exemplified by Uniswap v3, require liquidity providers (LPs) to actively manage price ranges to remain competitive against passive strategies. This paper presents the Automated Intelligent AMM (AIAMM) framework, which applies three deep reinforcement learning (DRL) architectures-Double Deep Q-Network (DDQN), Proximal Policy Optimization (PPO), and Soft Actor-Critic (SAC)-to the LP management problem. Under a rigorous IEEEcompliant experimental protocol (5 seeds × 500 episodes, curriculum learning, Bonferroni-corrected hypothesis testing), DDQN achieves statistically significant outperformance over a passive baseline: alpha = +$54.80 (95% CI: [41.19, 68.41]), Sortino ratio = 2.128, win rate = 41.2% at evaluation horizon T=500 (t = 7.896, padj < 0.001). PPO achieves borderline significance (padj = 0.009), while SAC remains non-significant. Earliest Effective Cutoff (EEC) bisection identifies T* ∈ (300, 500) steps, establishing the minimum horizon for positive expected alpha. These results constitute the first multi-architecture, multi-seed DRL benchmark for concentrated liquidity management with academically validated statistical inference.