Interpretable Markov-Based Spatiotemporal Risk Surfaces for Missing-Child Search Planning with Reinforcement Learning and LLM-Based Quality Assurance

The paper presents "Guardian," an interpretable, three-layer decision-support system that combines Markov chains, reinforcement learning, and LLM-based validation to generate dynamic, probabilistic search plans for missing-child investigations within the critical first 72 hours.

The Gist

Imagine you are trying to find a lost child in a massive, dark forest. You only know one thing: the exact spot where they were last seen. The clock is ticking, and every hour that passes makes the search harder.

This paper introduces Guardian, a high-tech "digital detective" designed to help police and search teams figure out where to look next. Instead of guessing, Guardian uses a three-step thinking process to turn messy, confusing reports into a clear, actionable map.

Here is how Guardian works, explained through simple analogies:

1. The Problem: The "Messy Pile"

When a child goes missing, investigators have a pile of clues: handwritten notes, PDF reports, text messages, and maps. It's like trying to solve a puzzle where the pieces are in different languages, some are torn, and some are just scribbles.

• Guardian's Job: It acts like a super-organized librarian. It takes all that messy information, cleans it up, and turns it into a structured list of facts (e.g., "Last seen at 3 AM," "Likely walking," "Near a highway").

2. The Three-Layer Brain

Guardian doesn't just guess; it uses a three-layer "brain" to predict where the child might be over the next 24, 48, or 72 hours.

Layer 1: The "Walking Simulator" (Markov Chain)

Think of this layer as a simulated ant walking across a giant map of the state.

• How it works: The ant starts at the "Last Seen" spot. But it doesn't just wander randomly. It follows rules:
  • Roads are highways: The ant prefers walking along roads because it's easier (lower cost).
  • Bushes are hiding spots: The ant is more likely to stop in quiet, secluded areas (high "seclusion score").
  • Day vs. Night: The ant behaves differently at night (maybe it hides more) than during the day.
• The Result: After simulating thousands of "ants" walking for 24, 48, and 72 hours, the system creates a heat map. Red areas mean "high chance the child is here," and blue areas mean "low chance."
• Key Feature: It knows that as time passes, the child could be further away, so the "red zone" gets bigger, but it doesn't spread everywhere equally. It follows the roads and terrain.

Layer 2: The "Resource Manager" (Reinforcement Learning)

Now we have a heat map, but search teams have limited people, dogs, and helicopters. They can't search the whole state at once.

• How it works: This layer acts like a smart chess player. It looks at the heat map and asks, "If I send a team to this specific square, how likely are we to find the child quickly?"
• The Goal: It draws circles and shapes on the map to create a search plan. It prioritizes the spots with the highest chance of success while avoiding overlapping teams (wasting effort). It tells the coordinators: "Send Team A to the woods here, and Team B to the highway exit there."

Layer 3: The "Human Sense-Checker" (LLM Quality Assurance)

Computers are great at math, but they can sometimes be weirdly logical. For example, a computer might suggest searching a spot that is mathematically probable but makes no sense in the real story (e.g., "The child is likely in a swamp, but the report said they were wearing nice shoes and wouldn't go there").

• How it works: This layer uses a Large Language Model (AI) that acts like a senior detective. It reads the original story and the computer's plan.
• The Check: It asks, "Does this plan make sense?" If the AI says, "Wait, this zone contradicts the witness story," it flags it or adjusts the priority. It ensures the plan isn't just mathematically correct, but also investigatively sensible.

3. The Test Run

The authors tested Guardian using a fake (but very realistic) case of a missing teenager in Virginia.

• The Result: The system successfully predicted that the child was most likely to be found in the Tidewater region (near the coast) within the first 24 hours, and then slowly spread toward Northern Virginia over the next few days, following the major highways.
• Why it matters: It showed that the system could narrow down a massive search area into specific, manageable zones, saving time and resources.

The Big Picture

Guardian isn't a robot that replaces human detectives. Instead, it's a super-powered assistant.

• It takes the chaos of a missing person case.
• It runs a simulation of how a person might move.
• It creates a strategy for where to look.
• It gets a second opinion from an AI to make sure the strategy makes sense.

By doing this, it helps search teams focus their energy on the most promising areas, giving them a better chance of finding a missing child in those critical first 72 hours.

Technical Summary

Here is a detailed technical summary of the paper "Interpretable Markov-Based Spatiotemporal Risk Surfaces for Missing-Child Search Planning with Reinforcement Learning and LLM-Based Quality Assurance."

1. Problem Statement

Missing-child investigations are time-critical, with the first 72 hours being decisive for recovery. However, law enforcement agencies face significant challenges:

• Data Fragmentation: Critical information is scattered across unstructured, heterogeneous sources (PDF reports, tips, maps, sensor data).
• Lack of Dynamic Tools: Traditional search planning relies on manual heuristics and human judgment, which struggle to synthesize complex geospatial and behavioral data under severe data sparsity.
• Operational Bottlenecks: The time required to convert fragmented narratives into a shared, geographically grounded search plan often delays resource allocation before the search area expands and uncertainty grows.

The goal is to create an end-to-end decision-support system that standardizes case evidence, generates probabilistic search surfaces, and produces actionable, interpretable search plans for 24, 48, and 72-hour horizons without replacing human investigative procedures.

2. Methodology: The Guardian System

The authors propose Guardian, a modular, two-stage pipeline designed to transform raw case documents into prioritized search zones. The system is built on a three-layer predictive architecture within its "Core System."

Stage 1: Data Pre-processing (Guardian Parser Pack)

• Input: Heterogeneous unstructured documents (e.g., NamUs, FBI, NCMEC reports).
• Process: A hybrid extraction pipeline using multi-engine text extraction (OCR + text engines) and source-aware parsing (rule-based templates for stable formats; LLM-assisted extraction for variable narratives).
• Output: Normalized, schema-aligned structured data (JSONL/CSV) enriched with geocoding, transportation context, and temporal metadata.

Stage 2: Analysis and Prediction (Guardian Core)

The core predictive engine consists of three distinct layers:

Layer 1: Markov Mobility Forecasting (The Predictive Core)

• Objective: Estimate the probability distribution of a missing person's location over time.
• Mechanism: Uses a sparse, interpretable Markov chain over a gridded geographical space (Virginia in the case study).
• Key Components:
  • Initialization: A mixture of a Gaussian seed (centered on the Last Known Position/IPP) and a historical hotspot prior (via Kernel Density Estimation).
  • Transition Matrices: Constructed using an energy-based approach incorporating:
    • Road Accessibility Costs: Lower cost for easier movement.
    • Seclusion Preferences: Higher probability for concealed terrain.
    • Corridor Bias: Proximity to major transport routes (e.g., interstates).
  • Day/Night Dynamics: Separate transition matrices for day and night to capture non-stationary mobility patterns (e.g., reduced travel at night).
  • Temporal Decay: A "survival-style" half-life decay ($2^{-t/T_{1/2}}$) is applied to widen uncertainty over time, preventing overconfidence in long-horizon predictions.
  • Boundary Masking: Enforces geographic constraints (e.g., state borders) to prevent probability leakage.

Layer 2: Reinforcement Learning (The Decision Layer)

• Objective: Convert the probabilistic belief maps from Layer 1 into a compact set of actionable search zones.
• Mechanism: Frames search planning as a sequential decision-making problem under resource constraints.
• Reward Function: Optimizes for:
  1. Early Capture: Maximizing probability mass covered early in the 0–72h window.
  2. Coverage Efficiency: Penalizing redundant overlap and excessive area.
  3. Plausibility: Ensuring zones align with corridor bias and seclusion preferences.
• Output: Ranked sectors, candidate search zones, and containment rings (50%, 75%, 90% quantiles) for specific time windows (0–24h, 24–48h, 48–72h).

Layer 3: LLM-Based Quality Assurance (The Validation Layer)

• Objective: Perform post-hoc validation to ensure search plans are operationally plausible and consistent with narrative constraints.
• Mechanism: Uses lightweight, instruction-tuned LLMs (Qwen-2.5-3B and LLaMA-3.2-3B) to review RL-generated zones.
• Process: The LLM analyzes the case summary, movement profile, and zone geometry to assign a plausibility score and a natural language rationale.
• Action: It re-weights the priority of zones (e.g., downgrading a mathematically optimal zone that contradicts a known behavioral cue) without altering the underlying Markov or RL models. This ensures the system remains auditable and interpretable.

3. Key Contributions

1. Three-Layer Architecture: A novel integration of Markov Models (for interpretable probabilistic forecasting), Reinforcement Learning (for resource-constrained optimization), and LLMs (for semantic quality assurance).
2. Interpretability: Unlike "black box" deep learning models, Guardian uses sparse Markov transitions and energy-based weights that can be explicitly traced to factors like road access or seclusion.
3. Human-in-the-Loop Design: The system is designed as a decision-support tool, not an autonomous authority. The LLM QA layer acts as a semantic sanity check to align statistical outputs with investigative reasoning.
4. End-to-End Pipeline: Successfully bridges the gap between unstructured PDF reports and structured, geospatial search products.

4. Results

The system was evaluated using a synthetic but operationally realistic case study (GRD-2025-001541) simulating a missing 15-year-old female in Virginia.

• Spatial Concentration: The model successfully identified the Tidewater region as the primary search area, retaining >50% of the total probability mass across all horizons. This was driven by the combination of the specific movement profile (on-foot, residential-local) and historical hotspot priors.
• Corridor vs. Seclusion: The model correctly balanced corridor bias (Northern Virginia emerged as a secondary region due to connectivity) against seclusion preferences.
• Temporal Evolution:
  • 24h: High concentration around the Initial Planning Point (IPP) with a 50% containment radius of ~20 miles.
  • 72h: Uncertainty expanded in a structured manner along transportation corridors rather than uniformly, with the 50% radius expanding to the mid-20s miles.
• Validation: The LLM QA layer successfully identified and re-weighted zones based on narrative consistency, demonstrating the value of the third layer in refining purely statistical outputs.
• Sensitivity: The system is sensitive to the mixing weight of priors ($\alpha_{prior}$), corridor/seclusion weights, and day/night switching schedules, highlighting the need for parameter calibration.

5. Significance

• Operational Impact: Guardian addresses the critical bottleneck of converting fragmented data into actionable search plans within the "golden 72 hours," potentially improving recovery rates.
• Responsible AI: By separating the predictive core from the decision layer and adding an LLM validation step, the system adheres to responsible AI principles for high-stakes humanitarian contexts. It provides auditable reasoning rather than opaque automation.
• Scalability: The modular architecture allows the system to be adapted for other populations (e.g., elderly missing persons) by calibrating movement profiles without changing the core algorithm.
• Future Directions: The authors propose future work on learning parameters from real-world retrospective data (under privacy safeguards) and integrating higher-order Markov dynamics to capture non-memoryless movement behaviors.

In summary, the paper presents Guardian as a robust, interpretable framework that leverages the strengths of probabilistic modeling, optimization, and generative AI to enhance the efficiency and effectiveness of missing-child search operations.