Abstract

Personalized treatment regimes that account for individual characteristics offer substantial potential to minimize treatment risks and enhance patient outcomes. Existing methods for estimating individualized treatment rules in multi-arm settings often face limitations due to model misspecification. To overcome the challenges, we propose a novel conditional generative learning framework, CG-Learning, which employs a Wasserstein generative adversarial network to estimate optimal decision rules that minimize a specified risk measure in multi-armed treatment scenarios. By modeling the conditional distribution of rewards through conditional generative adversarial networks, our approach avoids restrictive structural assumptions. We establish key theoretical properties of the resulting decision rules, including nonasymptotic bounds on regret and mis-assignment probabilities. We show that these bounds can be tightened when the covariate space has a low Minkowski dimension. Through comprehensive simulations and an application to data from the AIDS Clinical Trials Group, we demonstrate that CG-Learning outperforms existing benchmarks.