HMM-DQN Framework for 2026 F1 Energy Strategy Under Partial Observability
A recent paper on arXiv has proposed a new two-part system for making decisions about energy use in the 2026 Formula 1 season. The updated rules require a 50/50 balance between internal combustion engines and batteries, permitting unlimited energy recovery and a driver-controlled Override Mode. Since optimal energy management depends on the unseen states of rival cars, this scenario is classified as a Partially Observable Stochastic Game, which can't be tackled with standard single-agent methods. The first part employs a 40-state Hidden Markov Model to gauge the likelihood of competitors' ERS charge levels and additional factors from telemetry data. The second part features a Deep Q-Network that leverages the HMM results to craft energy usage strategies, ensuring the framework's practicality.
Key facts
- 2026 Formula 1 technical regulations introduce 50/50 ICE/battery power split
- Unlimited regeneration and driver-controlled Override Mode are allowed
- Optimal energy deployment depends on hidden states of rival cars
- Problem is a Partially Observable Stochastic Game
- First layer: 40-state Hidden Markov Model (HMM) for inference
- HMM infers rival ERS charge level (H, M, L_harvest, L_derate), Override Mode, tyre degradation
- Six publicly observable telemetry signals are used as input
- Second layer: Deep Q-Network (DQN) policy using HMM belief state
Entities
Institutions
- arXiv