Beyond Scripted Actors: Behavioural Simulation for Human Judgement Under Uncertainty
Military judgement is tested when other actors respond differently from what a plan assumes. An ally may hesitate, a civilian organisation may lose trust, or an opponent may interpret an action in an unexpected way. Training for these situations requires simulated counterparts whose decisions respond coherently to changing circumstances, incomplete information and the consequences of earlier interactions.
This paper presents a behavioural simulation methodology for creating and evaluating these counterparts in military training and wargaming. Drawing on Seldon’s work, it represents actors through explicit information access, beliefs, memory, objectives and relationships. As a scenario unfolds, these states shape what each actor perceives, how it interprets events and which actions it takes.
An illustrative multi-domain exercise explores how disrupted communications and conflicting information affect coordination, trust and escalation. Controlled variations change what actors know, whom they trust and what they believe others intend. These variations create opportunities for participants to recognise misunderstanding, question assumptions and adapt their decisions, while giving instructors visibility into the simulated behaviour driving each challenge.
The proposed evaluation compares scripted actors, persona-prompted language-model agents and agents with persistent behavioural state within a common scenario. It examines decision sequences, consistency with each actor’s information and history, sensitivity to controlled changes, variation across repeated runs and subject-matter expert assessment. A central question is whether behavioural variation reflects the scenario and actor state, or merely inconsistent model output.
The methodological contribution is an evaluation framework connecting agent behaviour to defined exercise learning objectives. The paper sets out practical recommendations for scenario design, instructor oversight and after-action review, alongside the further evidence required to establish training effectiveness.
For military users and simulation developers, the intended outcome is a repeatable way to create credible, inspectable decision challenges that exercise human judgement under uncertainty.