Proof of Concept / Game-Theoretic Multi-Agent RL
Nash Equilibrium Discovery in 1L NSCLC via Multi-Agent Q-Learning
Two competing oncology brand teams simultaneously learn optimal commercial strategies via independent Q-learning. The engine simulates head-to-head competition across 9 strategy combinations — pricing tier, messaging approach, and channel investment — and discovers Nash equilibria where neither brand can improve by unilaterally changing strategy. Built entirely in-browser with no backend dependencies.
Disclaimer: This is a proof-of-concept using synthetic competitive dynamics. All brand names, market share figures, and strategy payoffs are illustrative. This tool does NOT represent actual competitive intelligence or commercial strategies of any pharmaceutical company. Not intended for actual business decisions.
Simulating competitive dynamics...
Training 300 episodes × 24 months × 2 agents
Each brand maintains its own Q-table mapping (own_share_bucket × opponent_last_action) → strategy value. Brands update simultaneously using Bellman equations — treating the opponent as part of the environment. ε-greedy exploration decays over 300 training episodes.
Each of the 81 strategy pair combinations is simulated for 100 rounds to estimate expected market share changes for both brands. The resulting 9×9 payoff matrix encodes how each strategy performs against every competitor response — the foundation for Nash equilibrium analysis.
Pure-strategy Nash equilibria found by checking: can either brand improve by deviating? A strategy pair (a*, b*) is Nash if payoffA(a*, b*) ≥ payoffA(a, b*) for all a, and similarly for brand B. Nash proximity scores quantify near-equilibrium states.