Adaptive Monitoring of Stochastic Fire Front Processes via Information-seeking Predictive Control

1. Problem Statement

The paper addresses the challenge of adaptive monitoring of a stochastic wildfire front using a mobile agent (e.g., a drone). The core objective is to plan the agent's trajectory over a rolling finite horizon to minimize the uncertainty in estimating the fire's evolution.

Key Challenges:

• Stochasticity: Wildfire evolution is inherently stochastic, driven by random environmental factors (wind direction, wind speed, fuel type) that vary spatially and temporally.
• Integration Gap: Existing methods often decouple sensing, estimation, and control, or rely on simplifying assumptions (e.g., linear-Gaussian dynamics) that do not hold for complex fire propagation.
• Non-Convexity & Nonlinearity: The problem involves nonlinear fire dynamics, set-valued observations (multiple fragmented detections per fire vertex), and a non-Gaussian noise profile, making standard optimal control techniques (like Kalman filtering or standard MPC) inapplicable.
• Active Sensing: The agent must actively choose where to move to gather the most informative data, balancing immediate uncertainty reduction with future information gain (non-myopic planning).

2. Methodology

The authors propose a unified framework integrating sensing, estimation, and control within a Stochastic Optimal Control (SOC) framework.

A. System Modeling

• Fire Front Dynamics: The fire front is modeled as a stochastic adaptation of a deterministic elliptical growth model. The front is represented by $N$ vertices. The dynamics are governed by nonlinear differential equations where the velocity of each vertex depends on local wind direction ($\theta$), wind speed ($w_s$), and fire spread rate ($r_f$). These environmental parameters are treated as random variables, introducing process noise.
• Agent Dynamics: The mobile agent is modeled as a point-mass with discrete-time deterministic dynamics.
• Sensing Model: The agent uses a camera with a finite circular sensing range. Due to image processing imperfections, a single fire vertex may generate multiple fragmented "blobs" (measurements). The number of detections follows a Poisson distribution, and the measurement noise is Gaussian. This results in a random-cardinality point pattern observation model.

B. Recursive Bayesian Estimation

To handle the random cardinality of measurements (multiple detections per vertex), the authors derive a recursive Bayesian estimator specifically for this class of stochastic nonlinear elliptical models.

• Likelihood Function: Instead of standard vector likelihoods, they derive a set-likelihood function based on Poisson point processes. This allows the filter to handle multiple objects and multiple measurements without requiring explicit data association (matching specific measurements to specific vertices).
• Implementation: Since the system is nonlinear and non-Gaussian, the belief state is approximated using a Sequential Importance Resampling (SIR) Particle Filter.

C. Information-Seeking Predictive Control

The control problem is formulated as a finite-horizon Markov Decision Process (MDP) where the goal is to minimize the expected cumulative uncertainty (Risk-Weighted Dispersion) of the fire front belief over the horizon.

• Cost Function: The cost is defined as the Risk-Weighted Dispersion (RWD), which calculates the determinant of the covariance matrix of the fire front particles within high-risk grid cells.
• Policy Search (LCB-Guided): Since the state space (continuous belief space) is infinite and the problem is non-convex, standard Dynamic Programming is infeasible. The authors reformulate the problem as a Multi-Armed Bandit problem where each possible control sequence (policy) is an "arm."
  • They employ a Lower Confidence Bound (LCB) based adaptive search algorithm.
  • The algorithm iteratively simulates policies (rollouts) to estimate their expected cost.
  • It balances exploration (trying under-simulated policies) and exploitation (selecting the best current policy) using an LCB score derived from the Chernoff-Hoeffding bound.
  • Theoretical Guarantee: The algorithm is proven to asymptotically converge to the optimal policy with a regret decay rate of $O(\frac{\ln n}{n})$.

3. Key Contributions

1. Unified Framework: A novel integration of sensing, estimation, and control for stochastic fire monitoring, moving beyond decoupled or heuristic approaches.
2. Recursive Bayesian Estimator: Derivation of an optimal estimator for elliptical fire fronts with Poisson-distributed, random-cardinality measurements, eliminating the need for explicit data association.
3. LCB-Based Adaptive Control: Transformation of the complex nonlinear stochastic control problem into an MDP solvable via an LCB-guided search, providing asymptotic optimality guarantees without requiring linear-Gaussian assumptions.
4. Risk-Aware Planning: The cost function explicitly incorporates spatial risk maps, ensuring the agent prioritizes monitoring high-severity areas.

4. Evaluation and Results

The approach was evaluated in a simulation environment mimicking a 3km $\times$ 3km area with stochastic wind and fuel conditions.

• Setup: A drone with a 430m sensing range monitored a fire front represented by 20 vertices. The simulation ran for 25 time steps.
• Baselines: Compared against:
  • Infinite Sensing Range (Static): The theoretical lower bound (best possible performance).
  • Static Agent: Finite range, no movement.
  • Random Agent: Finite range, random movement.
  • Myopic Control: Non-myopic planning with horizon $T=1$.
  • Non-Myopic: Planning with horizons $T=3$ and $T=5$.
• Results:
  • Trajectory: The non-myopic approach ($T=3, 5$) successfully planned trajectories to target high-risk, high-uncertainty regions that the myopic agent ignored.
  • Accuracy: The proposed method significantly outperformed the Static and Random baselines.
  • Horizon Impact: Increasing the planning horizon ($T$) from 1 (myopic) to 3 and 5 resulted in a clear reduction in Root Mean Square Error (RMSE) of the fire front estimation. The non-myopic agent achieved estimation accuracy much closer to the theoretical "Infinite Sensing" baseline than the myopic agent.

5. Significance

This work provides a rigorous mathematical foundation for active wildfire monitoring in highly uncertain environments. By proving that an information-seeking control law can asymptotically converge to the optimal policy without relying on linear-Gaussian approximations, the paper offers a robust solution for real-world disaster response. The ability to handle complex, non-Gaussian sensor data (fragmented detections) and integrate it with stochastic fire dynamics makes this approach particularly valuable for deploying autonomous drones in dynamic, high-stakes scenarios like the 2025 Southern California wildfires mentioned in the introduction.