Strategic Trade Protection Under Incomplete Information: A Bayesian Learning Approach to Tariff Wars

This research pioneers a revolutionary synthesis between Bayesian game theory and deep reinforcement learning to decode the dynamics of strategic trade policy under incomplete information. We introduce the Logit-Quantal Response Equilibrium Bayesian (LQREB) framework—a novel computational architecture that bridges the theoretical elegance of game-theoretic equilibria with the adaptive power of neural networks. Through 10 000 simulated episodes modeling U.S.-Mexico-Canada trade interactions, the framework reveals emergent patterns of moderation: aggressive protectionist players (U.S.) converge to moderate tariffs (10%) when facing credible retaliation threats, vulnerable partners (Mexico) adopt passive strategies (97% non-retaliation), while strong partners (Canada) maintain firm responses (99% with 25% retaliation). The LQREB’S unique contribution lies in its ability to capture bounded rationality through temperature-controlled softmax (λ=20) while preserving Bayesian belief updating, achieving convergence where traditional Perfect Bayesian Equilibrium methods fail computationally. This work establishes the first scalable and falsifiable framework for analyzing multi-agent strategic protection under radical uncertainty.