Effective personnel training and sub-system testing in synthetic warfare environments has achieved incredible levels of realism and fidelity in physics and human-machine interface. However, progress has lagged in constructive behaviors—specifically, in planning and decision-making over long time horizons—with simulations still relying on scripted or rules-based behavior models. In contrast, we hypothesize that “intelligent” constructive behaviors can be realized automatically through maximization of long-term rewards. To test this, we design an implicit approach using Neural Network informed search, trained through self-play without input from a human expert, to estimate the best action for an entity and expected outcome of a scenario. We demonstrate emergent “intelligent” human-level behavior in multi-domain environments with electromagnetic actions and non-trivial collections of entities, with applications to adversarial training simulations.

Keywords
AGENT-BASED SIMULATION,BEHAVIOR MODELING,CONSTRUCTIVE,ELECTRONIC WARFARE,MACHINE LEARNING,PERSONALIZED TRAINING

Additional Keywords