Opponent State Inference Under Partial Observability: An HMM-POMDP Framework for 2026 Formula 1 Energy Strategy

Imagine Formula 1 racing in 2026 isn't just about who has the fastest car, but who is the best poker player.

In the old days, drivers knew exactly how much fuel and energy their rivals had left. It was like playing chess where you could see everyone's cards. But the new 2026 rules have changed the game. Now, every car has a massive battery that needs constant charging and draining, and drivers can choose to "go fast" (burn energy) or "go slow to charge" (harvest energy).

The problem? You can't see your rival's battery. You only see their speed, their braking, and where they are on the track. You have to guess what they are doing inside their car.

This paper is a guide on how to build a super-smart detective system (an AI) that can guess what your rival is thinking and plan your next move accordingly.

Here is the breakdown of how it works, using simple analogies:

1. The Core Problem: The "Blind Date"

Imagine you are on a date, but your partner is wearing a mask. You can see if they are smiling or frowning (observable data), but you don't know if they are actually happy, bored, or secretly planning to run away (hidden state).

In 2026 F1, a rival driver might be:

Full of energy: Ready to attack you.
Running on empty: Forced to go slow because their battery is dead.
Playing a trick: Pretending to be weak (going slow) to lure you into attacking, only to reveal they actually have a full battery and crush you.

This paper calls the third scenario the "Counter-Harvest Trap." It's a deceptive move where a driver hides their strength to trick you into wasting your own energy.

2. The Solution: Two Layers of Intelligence

The authors built a two-part AI system to solve this mystery.

Layer 1: The Detective (The HMM)

Think of this as a Sherlock Holmes inside the computer.

The Clues: It watches six specific things: speed differences, braking distance, tire wear, and a new "throttle signature" (how much the driver is pressing the gas pedal while the car is actually slowing down).
The Secret: It uses these clues to guess the rival's hidden state.
The Big Upgrade (v2): In previous versions, the AI just knew the rival was "Low on Energy." Now, it can tell the difference between two types of "Low":
- L-Harvest (The Trap): The driver is choosing to go slow to save energy. They are setting a trap.
- L-Derate (The Weakness): The driver has to go slow because their battery is physically dead. They are vulnerable.
Why it matters: If the AI thinks it's a trap, it won't attack. If it thinks it's a weakness, it will attack. This distinction is the paper's biggest breakthrough.

Layer 2: The General (The DQN)

Think of this as the Team Principal making the final call.

It takes the Detective's report (e.g., "There is an 89% chance the rival is setting a trap") and decides what to do.
The Choice: Should I "Burn" (use my energy to go fast and overtake) or "Harvest" (save my energy and wait)?
The General uses a neural network (a type of AI brain) that has learned from thousands of simulated races to make the perfect decision based on the Detective's guess.

3. The "Super-Clipping" Clue

The paper introduces a new "sixth sense" called $\delta$ throttle.

Imagine a car is going slow, but the driver is pressing the gas pedal 100% to the floor.
Scenario A: The driver is trying to go slow to save energy (The Trap).
Scenario B: The driver is pressing the gas, but the car won't go faster because the battery is dead (The Weakness).
The new AI can spot this tiny difference in the data. It's like hearing the difference between someone pretending to be tired and someone who is actually exhausted.

4. How They Tested It

Since the 2026 season hasn't happened yet (the paper is dated March 2026), they couldn't test it on real races. Instead, they created a video game simulation.

They built a virtual world with 40 different "moods" for the rival cars.
They ran 20 fake races.
The Result: The AI was incredibly good. It guessed the rival's battery level correctly 96.8% of the time (random guessing would only be 25%). It could spot the "Trap" 96.3% of the time.

5. The Catch (Limitations)

The paper admits one big flaw: The AI assumes the rival is a robot.

The current system assumes the rival driver doesn't know you are watching them.
In reality, if a rival driver realizes you have a super-smart AI, they might change their strategy to trick your AI.
The authors say, "This is just the first step. The next paper will teach the AI how to play against a rival who is also trying to trick it."

Summary

This paper presents a two-step AI strategy for 2026 F1:

Detect: Use math to guess if a rival is "faking it" (setting a trap) or "actually broken" (running out of power).
Decide: Use that guess to decide whether to attack or hold back.

It turns Formula 1 from a game of pure speed into a game of psychological warfare, where the winner is the one who can best read the mind of their opponent.

Here is a detailed technical summary of the paper "Opponent State Inference Under Partial Observability: An HMM–POMDP Framework for 2026 Formula 1 Energy Strategy" by Kalliopi Kleisarchaki.

1. Problem Statement

The paper addresses the strategic complexity introduced by the 2026 Formula 1 technical regulations, which fundamentally alter energy management. Key regulatory changes include:

50/50 Power Split: A mandatory split between the Internal Combustion Engine (ICE) and the MGU-K battery, placing the battery under continuous demand.
Override Mode: A proximity-gated energy award (0.5 MJ) replacing DRS, triggered when a car is within 1 second of the leader.
Active Aero: A driver-controlled binary aero system replacing DRS, which can be deployed in straight lines to gain speed independent of battery usage.

The Core Challenge:
Optimal energy deployment (Burn vs. Harvest) is no longer a single-agent optimization problem. It depends on the hidden state of rival cars (battery charge, Override Mode status, and tire degradation). This creates a Partially Observable Stochastic Game (POSG).

The Counter-Harvest Trap: A specific deceptive strategy where a leading car (Car B) deliberately conserves energy (Lharvest mode) while deploying Active Aero to maintain speed. This tricks the following car (Car A) into believing Car B is physically depleted (Lderate), prompting Car A to burn its own energy to attack. Car B then reveals its hidden reserve to defend easily.
Observability Gap: Drivers and teams only have access to public telemetry (speed, sector times, braking points), not the internal battery state or strategic intent of rivals.

2. Methodology

The authors propose a two-layer inference and decision framework approximating the POSG as a single-agent Partially Observable Markov Decision Process (POMDP).

Layer 1: Rival State Inference (Hidden Markov Model - HMM)

The first layer infers the probability distribution of a rival's hidden state based on six observable telemetry signals.

State Space (40 States): The hidden state $x_t = (e, m, \tau)$ $x_{t} = (e, m, τ)$ is decomposed into:
- ERS Charge ( $e$ ): 4 modes: H (High), M (Medium), Lharvest (Deliberate conservation/trap), Lderate (Physical depletion/SOC ceiling).
- Override Mode ( $m$ ): Available or Spent.
- Tire Degradation ( $\tau$ ): 5 levels (New to Cliff).
Observables (6 Signals):
1. $\Delta v_{trap}$ : Speed trap deviation.
2. $\Delta t_{sector}$ : Sector time deviation.
3. $\Delta b_{brake}$ : Braking point deviation.
4. $\sigma^2_{speed}$ : Speed trace variance.
5. $z_{aero}$ : Active Aero deployment status (binary).
6. $\delta_{throttle}$ (New in v1.5/v2): The fraction of time the throttle is $\ge 98\%$ while speed is below the rolling baseline. This is the critical discriminator between Lharvest (low $\delta_{throttle}$ , managed throttle) and Lderate (high $\delta_{throttle}$ , full throttle but power capped).
Inference Mechanism:
- Uses a 40-state HMM with a transition matrix derived from physics and an emission matrix derived from Gaussian/Bernoulli distributions.
- Calibration: Parameters are initialized analytically and refined using the Baum-Welch algorithm on real race telemetry (starting with the 2026 Australian GP).
- Architectural Advance (v2): Unlike previous versions where Lharvest and Lderate were a single "Low" state with mixed emissions, v2 treats them as distinct hidden states. This allows the belief state to explicitly output $P(Lharvest)$ and $P(Lderate)$ as separate dimensions.

Layer 2: Decision Policy (Deep Q-Network - DQN)

The second layer selects the optimal action (Burn or Harvest) based on the HMM's belief state.

Input: A 66-dimensional vector comprising the 40-dimensional HMM belief state, own-car state (ERS, fuel, tires, gaps), and race context.
Architecture: A shallow Deep Q-Network (3 hidden layers: 256-256-128) trained via Double DQN with experience replay.
Training: Pre-trained on synthetic data generated from the parametric model, then fine-tuned on real telemetry.
Reward Function: Maximizes race position change over a 5-lap horizon, with shaping rewards to penalize falling into the "counter-harvest trap."

3. Key Contributions

Problem Formalization: Models 2026 F1 energy strategy as a POMDP and formally defines the Counter-Harvest Trap as a deceptive equilibrium strategy.
State Decomposition (The v2 Advance): Introduces the $L_{harvest}$ vs. $L_{derate}$ split. This resolves the ambiguity where a car slowing down could be either "trapping" (strategic) or "depleted" (physical). The inclusion of $\delta_{throttle}$ as a state-level discriminator (rather than just an emission feature) is the central architectural innovation.
Tractable Inference Framework: Demonstrates a working pipeline combining HMM-based belief updates with DQN-based policy selection, capable of detecting traps that threshold-based baselines miss.

4. Results (Synthetic Validation)

The system was evaluated on 20 synthetic races generated under the 40-state model.

ERS Level Accuracy: 96.8% (Random baseline: 25%).
$L_{harvest}$ vs. $L_{derate}$ Classification: 89.4% accuracy (Random baseline: 50%). This confirms the efficacy of the $\delta_{throttle}$ signal.
Trap Detection Recall: 96.3% (False positive rate significantly reduced compared to v1.5).
Gap to Oracle: The system achieves 95.7% of the performance of an "Oracle" policy that knows the true hidden state.
Robustness: The model remains robust to parameter misspecification (e.g., mean shifts in emission distributions), though performance degrades slightly under high noise.

5. Significance and Limitations

Significance:

Strategic Insight: The paper proves that detecting the "Counter-Harvest Trap" is impossible with single-observable thresholds (like speed alone) but is solvable via multi-modal belief inference.
Regulatory Adaptation: Provides a concrete computational framework for teams to adapt to the 2026 regulations, moving from deterministic pit-stop optimization to dynamic, opponent-aware energy management.
Foundation for Game Theory: The v2 framework serves as the necessary "floor" (baseline) for future work on non-stationary, game-theoretic opponents (where rivals adapt to being observed), which is the subject of the author's forthcoming work.

Limitations:

Stationary Opponent Assumption: The current model assumes rivals do not adapt their strategy based on being observed. In reality, a rational opponent might change tactics if they know they are being inferred upon.
Conditional Independence: The HMM assumes observables are conditionally independent, which is physically untrue (e.g., speed and sector time are coupled). This leads to over-confident beliefs (low Expected Calibration Error on synthetic data, but likely higher on real data).
Data Scarcity: The model relies on Baum-Welch calibration. The authors note that meaningful empirical validation requires data from at least 4 races (approx. 14,000 sector observations) to overcome the parameter-to-data ratio constraints.
Exclusion of Boost Mode: The model currently excludes the "Boost Mode" (unconstrained deployment) as a latent dimension, focusing solely on the strategic interaction of the battery and Override Mode.

Conclusion

This paper presents a rigorous, data-driven framework for solving the partial observability problem in 2026 F1. By elevating the distinction between "strategic conservation" and "physical depletion" to the state level, the proposed HMM-DQN architecture successfully identifies deceptive strategies that would otherwise cause a team to lose a race. It establishes the baseline for future research into fully adversarial, game-theoretic racing strategies.