Bridging game-theoretic modeling and Artificial Intelligence for fast policy recommendations in multi-channel retail

Emerging-market milk retail is increasingly split across General Trade, Modern Trade, and E-commerce, where pricing and advertising are strategically interdependent through cross-channel substitution, advertising spillovers, and trust-sensitive demand. Yet decision support is challenging: game-theoretic models capture these interactions but are costly to solve repeatedly, while standard machine learning approaches typically require rich quantitative data, often unavailable in fragmented markets where key drivers (e.g., “low trust”) are qualitative. This paper bridges this gap by proposing a hybrid simulation–surrogate decision-support framework that operationalizes omnichannel theory into rapid decision support. First, a tri-channel game-theoretic simulator incorporating trust, advertising spillovers, and cost heterogeneity generates constrained solver benchmark pricing and advertising decisions. Second, to reduce repeated equilibrium-solving costs during scenario evaluation, a lightweight policy network is trained to approximate this structural mapping, providing feasible scenario-based recommendations at low computational cost. A rule-based natural-language descriptor layer operationalizes qualitative managerial inputs by mapping predefined cues into parameterized intervals. Across 10,000 simulated market states, the trained surrogate policy closely approximates the solver-generated benchmark decisions, with evaluation on the held-out test set yielding a mean profit ratio of 1.001, a trimmed profit ratio of 0.999, and mean economic regret of 0.40%. Stress tests under alternative parameter distributions and comparative evaluations against linear, tree-based, and heuristic baselines indicate the network's stable economic performance across regular simulated market states; while batched evaluations accelerate interactive scenario analysis by up to 14.8 times.