PaperlayerView exact revision on arXiv

Source-linked AI summary

PoKER: A Probability of Kill Estimation Rate Model for Air-to-Air Missiles Using Machine Learning on Stochastic Targets

Joao P. A. Dantas, Andre N. Costa, Diego Geraldo, Marcos R. O. A. Maximo, Takashi Yoneyama

arXiv:2610.32864v1cs.LGcs.AI

TL;DR

WEZ ranges may not give pilots a clear launch choice when an opponent maneuvers unpredictably. PoKER combines missile simulation, probabilistic target behavior, and warhead lethality to estimate kill probability; ANN predicts better than PR without a significant increase in training time.

Abstract

from arXiv · show

This work introduces PoKER, a novel probabilistic model engineered to optimize missile launch effectiveness in air-to-air scenarios, specifically within Beyond Visual Range (BVR) air combat. Unlike conventional Weapon Engagement Zone (WEZ) models that delineate zones based on static distances such as maximum, minimum, and no-escape ranges, PoKER applies machine learning to predict kill probabilities more accurately by integrating the stochastic behaviors of targets and missile miss distances. This model dynamically evaluates target behavior, greatly expanding the predictive capabilities of engagement analysis. By factoring in elements such as warhead lethality, target and shooter orientations, and the specific conditions of engagement, PoKER provides important insights into engagement dynamics and quantifies success probabilities. Consequently, it can potentially be an important tool for BVR air combat pilots, improving operational decision-making within this specialized combat domain.

Introduction

Traditional WEZ analysis uses engagement ranges and orientations, but may leave pilots without a clear launch choice when targets maneuver unpredictably. PoKER combines simulation, target behavior, and warhead lethality with machine learning to estimate kill probability in dynamic BVR engagements.

Introduction

  • Unpredictable target maneuvers and broad engagement regions can make it difficult for pilots to choose a launch moment that maximizes success.WEZ terminology includes maximum ranges, no-escape regions, and the minimum arming distance.Source: Introduction · pp. 1–2
  • PoKER complements traditional WEZ analysis with kill-probability estimates that incorporate real-time launcher and target conditions.The model is designed to improve situational awareness in dynamic BVR combat.Source: Introduction · p. 2
  • PoKER integrates potential adversary maneuvers through high-fidelity aircraft and missile models rather than relying on fixed target profiles and launcher strategies.The study presents an R-based simulator using a 5 Degrees of Freedom model for an Active Radar Homing weapon system.Source: Introduction · p. 2
  • The approach combines simulation and supervised machine learning with a warhead lethality model based on missile miss distance and dynamic opponent behavior.The study evaluates algorithms for predictive performance and computational efficiency.Source: Introduction · p. 2
  • The authors present PoKER as an initial step toward probabilistic weapon-engagement analysis that could support more informed tactical decisions in complex air combat.The model is intended to account for dynamic interactions between launcher and target.Source: Introduction · p. 2

Related Work

Related work covers air-combat simulation, methods for defining engagement and escape zones, and cooperative tactics involving multiple aircraft. Existing approaches include simulated missile flyouts, lookup tables, and optimization of group engagements.

Related Work

  • Prior air-combat simulation studies examine weapon engagement, evasion strategies, and cooperative tactics to characterize dynamic combat scenarios [12–16].Source: Related Work · p. 3

Engagement and Escape Zone Strategies

  • WEZs support launch and guidance decisions through boundaries derived from ongoing flyout simulation or offline simulated data with interpolation [17–21].Related methods also use offline simulation, polynomial, interpolation, or neural-network fitting algorithms [4, 22–26].Source: Engagement and Escape Zone Strategies · p. 3

Cooperative Engagement Models

Probabilistic Modeling in BVR Air Combat

The paper extends WEZ analysis by modeling target maneuver probability and warhead lethality to estimate engagement success, using a missile simulator and supervised learning. It describes the simulator’s target, trajectory, flight-dynamics, and seeker behaviors, with ASA used for validation and debugging.

Limitations in Existing Work

Methodology

  • The methodology generates and preprocesses simulation data, calculates warhead lethality from miss distance, and trains supervised regression models to predict kill probability.It also estimates turn degree and delay as factors in computing the probability of hitting a target.Source: Methodology · p. 4

Missile Launch Simulator

  • The simulator models a Fox 3 ARH missile with a 5DOF framework, and its missile model is restricted due to confidentiality and sovereignty concerns.The simulator is implemented in R and is based primarily on prior work.Source: Missile Launch Simulator · p. 4

Guidance and Navigation

  • Proportional navigation guides the missile toward the moving target, while an optional loft maneuver sends it upward after launch.The passage describes lofting as critical to extending range and enhancing interception probability for distant targets.Source: Guidance and Navigation · p. 4

Target Interaction

  • Target simulations include stationary and maneuvering cases, including +5 G maneuvers whose timing can vary across engagement scenarios.Sample target metrics include velocity, acceleration, altitude, heading, and pitch.Source: Target Interaction · p. 4

Missile Trajectory Metrics

  • Missile trajectory metrics show near-linear mass decrease during boost and pronounced angle-of-attack and pitch variations during complex maneuvers such as lofting.Figure 4 presents these performance metrics across the simulation.Source: Missile Trajectory Metrics · pp. 4–5

Flight Dynamics

  • The flight-dynamics simulation tracks heading adjustments, North-East accelerations, and velocity that rises during boost and gradually falls during sustain.These metrics are used to assess responses to loft maneuvers and target evasion.Source: Flight Dynamics · p. 5

Seeker Behavior

  • Seeker behavior is assessed through seeker angle, which reflects adjustments in following a proportional-navigation path and deviations from off-boresight angle.The passage connects these deviations to assessing effectiveness against evasive targets.Source: Seeker Behavior · p. 5

Simulation Data

Simulation data are generated by sampling engagement conditions and running a high-volume missile simulator. The inputs describe shooter and target conditions, while outputs include missile miss distance.

Simulation Data

  • LHS sampled across prespecified sections of the input space, and military fighter pilots established parameter ranges as subject matter experts.The input intervals are presented in Table 1.Source: Simulation Data · p. 5, Table 1 · p. 5
  • Inputs include shooter altitude, velocity, and pitch; target velocity, heading, and off-boresight angle; range; and target response and maneuver parameters.The parameters capture engagement geometry and conditions affecting missile energy, flight time, and the target’s defensive maneuver.Source: Simulation Data · p. 5, Figure 5 · p. 6
  • 10 million LHS-generated input files produced 10 million missile-launch simulations, each recording miss distance at the target’s closest point of approach.Runs took approximately 40 days on an Intel Xeon Silver 4210R CPU with 128 GB RAM.Source: Simulation Data · p. 6
  • The model’s idealized shooter and target behavior omits factors such as pilot decision-making under uncertainty, weather, and missile degradation.The authors caution that parameter ranges and resolutions may not cover all operational scenarios, limiting generalizability.Source: Simulation Data · p. 7

Preprocessing

Preprocessing identifies correlated or improbable samples, while a physics-based warhead model converts detonation conditions and miss distance into a probability of kill. The model combines fragment behavior and lethality under stated assumptions.

Preprocessing

  • Target turn degree and acceleration were substantially correlated, so the authors proposed excluding load factor from machine-learning inputs.Exploratory analysis used descriptive statistics, histograms, box plots, correlations, and outlier analysis.Source: Preprocessing · p. 7

Warhead Lethality

  • The warhead model maps detonation conditions and relative target position to Pkill, which decreases as detonation distance moves away from the lethal radius.The probability depends on fragment velocity, fragment dispersion density, and miss distance.Source: Warhead Lethality · p. 7, Figure 6 · p. 7

Factors Affecting Warhead Lethality

  • Lethality depends on miss distance, fragment distribution and density, fragment initial velocity, and warhead and explosive properties.The study models pre-formed fragments in a brickwork pattern.Source: Factors Affecting Warhead Lethality · pp. 7–8

Modeling the Physical Properties of the Warhead

Warhead Performance Metrics

  • The Gurney equation estimates initial fragment velocity from the explosive’s Gurney constant and the explosive-to-total-warhead-mass ratio.Static ejection angles specify the fragment dispersion pattern relative to the warhead.Source: Warhead Performance Metrics · p. 8

Endgame Condition Determination

Endgame Dynamics Computation

  • Dynamic ejection angles account for missile velocity at detonation, and relative fragment-target velocity is used to estimate fragment travel and impact potential.Source: Endgame Dynamics Computation · p. 9

Computation of Final Probability of Kill

Supervised Machine Learning Models

The authors train regression models to predict Pkill calculated from simulated miss distance. They compare polynomial regression, artificial neural networks, and XGBoost using stratified train-validation-test splits.

Supervised Machine Learning Models

  • Polynomial regression, artificial neural networks, and XGBoost were trained to predict Pkill derived from simulated miss distance.The input set excludes load factor; turn degree and response delay require assumptions because they are not directly available to the launcher.Source: Supervised Machine Learning Models · p. 9
  • Data were stratified into 34 probability bins, with 80% assigned to training and validation using five-fold cross-validation and 20% to testing.Source: Supervised Machine Learning Models · p. 9
  • Polynomial regression degrees one through four provide a simpler baseline, while ANN and XGBoost represent more complex model options.The authors selected the three approaches to compare simplicity and performance.Source: Supervised Machine Learning Models · pp. 9–10
  • ANN configurations varied hidden layers and units, used scaled inputs and Adam optimization, and applied early stopping with ten-epoch patience.XGBoost hyperparameter exploration included tree depth, learning rate, and sampling rates, with early stopping also used.Source: Supervised Machine Learning Models · p. 10

Models Evaluation

The evaluation compares regression accuracy and computational cost, including inference time as a measure relevant to real-time applicability.

Models Evaluation

  • Models are evaluated using MAE, MSE, RMSE, and R², alongside training time and test-set inference time.The authors emphasize inference time because it directly affects practical applicability in real-time scenarios.Source: Models Evaluation · p. 10

Estimating Turn Degree and Delay

PoKER estimates target turn degree and reaction delay using SME-informed scenarios to represent evasive behavior and its variability. The estimates are face-validated by SMEs.

Estimating Turn Degree and Delay

  • SMEs set normal-distribution medians and standard deviations for target turn degree and delay to represent variable responses across scenarios.The scenarios focus on turn degree and the delay before the target begins turning.Source: Estimating Turn Degree and Delay · p. 11
  • Targets are modeled as taking the most defensive feasible turn, up to 180°, to maximize separation from the launcher.The initial turn degree varies with the target’s relative position to the launcher.Source: Estimating Turn Degree and Delay · p. 11
  • The final turn degree adds 5% variation using a standard normal random variable around the value calculated from Equation 2.This variation is intended to account for natural fluctuations in the target’s maneuver.Source: Estimating Turn Degree and Delay · p. 11
  • Reaction-delay assumptions span 15–30 seconds across three 5-second intervals, assigned to 120° orientation ranges covering 360°.The median and standard deviation estimates were face-validated by SMEs using simulated scenarios with varied distances, maneuvers, and launches.Source: Estimating Turn Degree and Delay · pp. 11–12

The Probability of Hitting a Target

PoKER estimates hit probability by incorporating inferred target turn degree and delay into machine-learning predictions. The estimates account for evasive behavior, but further empirical validation is needed to demonstrate their practical benefits.

The Probability of Hitting a Target

  • Because target turn degree and delay are not directly observable, PoKER estimates them for use as inputs to machine-learning hit-probability predictions.The estimates represent variability in target response behavior.Source: The Probability of Hitting a Target · p. 12
  • The model assumes high target performance and conservative evasive estimates to reflect challenging engagement scenarios in hit-probability predictions.The stated aim is to avoid underestimating the target’s evasive capabilities.Source: The Probability of Hitting a Target · p. 12
  • Further empirical validation is required to demonstrate the practical benefits of incorporating turn-degree and delay estimates.The authors describe the current work as a foundation for future assessment of these estimates’ impact on aerial defense strategies.Source: The Probability of Hitting a Target · p. 12

Results and Analysis

The results section examines test-dataset metrics and a display combining traditional range indicators with PoKER’s probability-of-kill indication. Model comparisons report predictive-error, R², and computational-efficiency measures.

Results and Analysis

  • The test-dataset results include an MFD representation combining maximum-range and no-escape-zone indicators with PoKER’s probability-of-kill indication.The section also examines test-dataset metrics for the proposed models.Source: Results and Analysis · p. 12

Comparative Analysis of the Machine Learning Models

  • The model comparison reports MAE, MSE, RMSE, and R² alongside training and inference times for PR, ANN, and XGBoost.Metrics summarize five folds using each model’s best validation-selected hyperparameter settings; Pkill is expressed as a percentage.Source: Comparative Analysis of the Machine… · p. 12, Table 3 · p. 12

Statistical Analysis

Across the tested models, ANN has better predictive accuracy than PR, while XGBoost’s results fall between them without significant pairwise differences from either. Statistical tests also identify training-time differences, while model complexity involves an accuracy–inference-time balance.

Friedman Test

  • The Friedman test found significant differences among models for MAE, MSE, RMSE, R², and training time, but not inference time.The authors note that five folds and correlated error metrics may limit detection of finer performance differences.Source: Discussion · p. 13, Friedman Test · p. 13

Nemenyi Post-hoc Test

  • Nemenyi tests found ANN significantly outperformed PR on MAE, MSE, RMSE, and R² (p-value = 0.0045), while XGBoost differed significantly from neither.XGBoost comparisons with PR and ANN each had p-value = 0.2541 across the reported accuracy metrics.Source: Discussion · pp. 13–14

Discussion

  • ANN achieved the lowest MAE (2.33), MSE (51.90), and RMSE (7.20), and an R² of 94.38%; PR had the highest errors and lowest R².PR’s MAE was 11.52, MSE 296.26, RMSE 17.21, and R² 67.94%.Source: Discussion · pp. 13–14
  • PR’s training time (532.23 seconds) was significantly shorter than XGBoost’s (6098.33 seconds), while ANN’s 4722.17 seconds differed significantly from neither.ANN’s inference time was 6.65 seconds, compared with 8.20 seconds for XGBoost and 9.57 seconds for PR.Source: Discussion · p. 13
  • More complex neural-network configurations increased inference time, while unit and layer choices affected mean R², requiring an accuracy–efficiency trade-off.The cited discussion describes an interplay between network depth and performance.Source: Discussion · p. 14
  • The authors caution that computational-time discrepancies may reflect different library implementations for PR and ANN.Their summary presents ANN’s training time as not significantly different from PR’s despite being longer.Source: Discussion · pp. 13–14

Enhanced Weapon Engagement Zone Representation

PoKER adds a probability-of-kill dimension to traditional WEZ displays, pairing it with maximum-range and no-escape indicators. The authors describe potential decision-support benefits while identifying validation, computational-scale, and reproducibility constraints.

Enhanced Weapon Engagement Zone Representation

Conclusion and Future Work

  • Future work proposes real training-launch data, human-in-the-loop simulations, and empirical studies across scenarios to refine target behavior modeling and validate effectiveness.Other proposed directions include adaptive machine learning, improved missile simulations, environmental and electronic-warfare factors, adversary tactics, and guidance toward higher-probability regions.Source: Conclusion and Future Work · pp. 14–15
  • The study used 10 million simulations, which may not be practical for all scenarios; the authors propose reducing dataset size while preserving model fidelity.Suggested approaches include efficient sampling or selecting representative points.Source: Conclusion and Future Work · p. 14

Supplemental Material

  • The methodology code is available, but classified data prevent full reproduction of the results.The shared code provides insight into the methods used.Source: Supplemental Material · p. 15

Author Biographies

The biographies describe researchers whose backgrounds span engineering, military research, artificial intelligence, and control theory.

Author Biographies

  • Joao P. A. Dantas and Andre N. Costa have engineering degrees and conduct Brazilian Air Force research involving artificial intelligence, machine learning, modeling, or simulation.Dantas is pursuing a Ph.D. at ITA; Costa’s stated interests include modeling and simulation.Source: Author Biographies · p. 18
  • Diego Geraldo is a fighter pilot and Brazilian Air Force researcher specializing in modeling, simulation, and artificial intelligence.He earned a master’s degree in Aeronautical and Mechanical Engineering from ITA.Source: Author Biographies · p. 18
  • Marcos R. O. A. Maximo is an ITA professor researching robotics and artificial intelligence, while Takashi Yoneyama is a control-theory professor focused on stochastic optimal control.Both biographies list academic appointments at ITA.Source: Author Biographies · p. 18

Related papers

Loading 2610.32864v1…