Q-Learning Model for Proportionality Assessment in Military Operations

Abstract

In the context of military operations, accurate and transparent proportionality assessment is essential to ensure
compliance with international humanitarian law. On this behalf, this research presents and evaluates two Q-learning
models designed to build the proportionality assessment in military operations. In this sense, the first model considers
collateral damage exclusively in physical terms (excluding psychological harm), while the second model explicitly integrates
psychological damage as part of the collateral damage effects. Both models encode operational rules as multi-attribute
states encompassing injury severity, fatalities, object damage, and military advantage, differing only in the inclusion of
psychological factors. From training and simulation results, it can be seen that this approach provides a valuable
classification approach for proportional and disproportional outcomes within their respective scenario sets. This shows
that AI (Artificial Intelligence) provides effective methods and techniques that allow both modelling and the expansion of
existing definitions and perspectives of existing challenging concepts in the uncertain and dynamic space of the military
domains while accounting and respecting legal and ethical considerations in order to build responsible and trustworthy
military AI systems.