Opponent State Inference Under Partial Observability: An HMM-POMDP Framework for 2026 Formula 1 Energy Strategy
New AI framework uses 30-state HMM and Deep Q-Network to solve 2026 F1's deceptive energy strategy game.
A new research paper by Kalliopi Kleisarchaki proposes a sophisticated AI framework designed to master the complex energy strategy game that will define the 2026 Formula 1 season. The upcoming technical regulations mandate a 50/50 power split between internal combustion engines and batteries, introducing unlimited energy regeneration and a driver-controlled Override Mode (MOM). This creates a Partially Observable Stochastic Game where a team's optimal energy deployment depends critically on the hidden battery charge and tire state of rival cars, a problem traditional single-agent optimization cannot solve. The paper, pre-registered on arXiv, presents a formal solution to this high-stakes strategic challenge.
The proposed framework is a two-layer system combining inference and decision-making. The first layer is a 30-state Hidden Markov Model (HMM) that processes five publicly observable telemetry signals to generate a probability distribution over each rival's Energy Recovery System (ERS) charge, Override Mode status, and tire degradation. This 'belief state' feeds into the second layer: a Deep Q-Network (DQN) policy that selects the optimal energy deployment strategy. Crucially, the research formally characterizes a 'counter-harvest trap'—a deceptive tactic where a car hides its true energy state to lure a rival into a failed overtake—and proves that detecting it requires this belief-state inference, not simple reactive rules. In synthetic race simulations, the HMM achieved 92.3% accuracy at inferring ERS state (vs. a 33.3% random baseline) and detected trap conditions with 95.7% recall. Empirical validation with real 2026 race telemetry is scheduled to begin at the season-opening Australian Grand Prix on March 8, 2026.
- The AI uses a 30-state Hidden Markov Model to infer rival car energy and tire states from public telemetry with 92.3% accuracy.
- It formally models and detects 'counter-harvest trap' deceptive strategies with 95.7% recall, requiring belief-state inference over simple rules.
- Real-world testing and calibration of the framework is pre-registered to begin with the 2026 F1 season at the Australian Grand Prix.
Why It Matters
Demonstrates how advanced AI can solve real-world, high-stakes strategic games with hidden information, moving beyond board games to professional sports.