SNAPO: Smooth Neural Adjoint Policy Optimization for Optimal Control via Differentiable Simulation

Battle Count: 50

Directly relevant to commodity trading (gas storage optimization with forward curve sensitivities/Greeks), portfolio optimization (pension ALM rebalancing), and risk management (sensitivity computation at scale). The adjoint-based Greeks computation (365 forward curve sensitivities at constant cost) is highly relevant for derivatives desks. However, the paper focuses on optimal control/decision-making rather than alpha generation or market prediction. The framework is more applicable to operational trading decisions (when to inject/withdraw) than to statistical arbitrage or high-frequency strategies.

Requires: (1) a known differentiable simulator with tape recording capability, (2) a compiled AD library (AADC or equivalent) for efficient adjoint passes, (3) careful design of smooth constraint relaxations with quantified bias, (4) Monte Carlo simulation infrastructure, (5) domain-specific physics/dynamics modeling. The neural policy itself is simple (small MLP), but the surrounding infrastructure (differentiable simulator, constraint smoothing, adjoint kernel compilation) is substantial. The commercial AADC dependency adds cost and limits open-source reproducibility.

Detailed hardware specs (Intel Xeon Platinum 8280L, AVX2, single-thread, no GPU), hyperparameters (lr=0.005, beta1=0.9, beta2=0.999, epsilon=1e-8), random seeds (0-4 for SNAPO, 0-2 for baselines), Monte Carlo paths, iteration counts, and validation protocols are all specified. Code and trained policies stated as 'available on request' but no public repository link provided. Appendix D includes reproduction commands. The AADC library is commercial, limiting full reproducibility.