Dr Rick Evertsz presented the paper "Adding Realistic Cognition/Emotion to VBS2 Entities" at the SimTecT 2009 Conference, Adelaide Convention Centre.
VBS2™ provides impressive photo-realistic 3D rendering for military training. Nevertheless, as is the case with all synthetic environments that focus on visual fidelity, the modelling of human behaviour in VBS2 is limited. The state of the art in human behaviour representation far exceeds what is currently embedded within synthetic environments such as VBS2.
Rick Evertsz reported on AOS's integration of the cognitive architecture CoJACK™ with VBS2. CoJACK is a novel cognitive architecture based on the BDI (Beliefs/Desires/Intentions) model that includes a tightly integrated moderator layer for modelling emotion and physical factors such as fatigue. One of CoJACK's discriminators is its ease of use. Human behaviour models are represented using a graphical plan language that is easily understood and verified by Subject Matter Experts (see diagram below).
The paper focused on a suicide bomber scenario in which fear and morale are implemented as CoJACK moderators that affect the agent’s perception of the utility of particular plans of action (e.g. fleeing).
VBS2 has a scripting language that can be accessed via the Application Scripting Interface (ASI) to allow control by external AIs. ASI provides a wide range of scripting commands, many suitable for human behaviour modelling.
The CoJACK/VBS2 Integration Layer maps incoming VBS2 messages to VBS2-independent Perception Events. VBS2 (on state of the art hardware) is able to render at a rate upwards of 70 frames per second, and can feed these updates at the same rate to CoJACK. This high-speed, discrete representation of environmental change is not how humans perceive an essentially continuous world. Therefore, CoJACK maps these to aggregate events that summarise a stream of percepts of a given type. This aggregation is a function of the type of data and can include an average, maximum or minimum value (or any user-defined aggregation function). These ephemeral aggregate events are then converted to memory elements and stored in Sensory Memory. Sensory Memory forms a snapshot of the current situation and is used to inform the next Situation Assessment phase of the process. Thus, perception is largely independent of the feed rate of VBS2.
In this approach, CoJACK represents the entity’s “mind” and is responsible for its decision-making and actions. VBS2 is the mediator of perception, action and environmental influences such as terrain and temperature. In effect, VBS2 is the representation of the entire physical world, apart from the internal state of the agent. The CoJACK agent maintains information about its own mental and physiological state, whereas VBS2 maintains information about everything outside the agent’s body.
The schematic shows percepts and actions being communicated between CoJACK and VBS2. Fear resides in the moderator layer. It takes input from the environment but can also take input from the cognitive model (e.g. a negative task appraisal can lead to an increase in fear).
A currently used and appropriate way to model fear is to represent it as a reservoir. When fear is triggered in the behaviour model, the amount of fear is used to initialize the level of the fear reservoir (instantaneously). The level of fear in the reservoir then decays over time with a half-life that is provided as an input to the moderator. This approach appears to be consistent with data on the effect of suppressive fire. This provides a robust implementation, and one that can be expanded and modified as additional data and needs arise. Variability between agents can be provided by varying the uptake and decay of fear constants as well as the base level. Variability with time is provided by the decay of fear. This approach is useful as a first step, as we want the models to be more complex but we also need them to not be too complex and difficult to understand, modify, and debug.
The development of CoJACK was funded by the IHBR (Improved Human Behaviour Representation) project, funded by DAES (Directorate of Analysis, Experimentation and Simulation) UK Ministry of Defence.
We would like to thank Bohemia Interactive for their assistance with the CoJACK/VBS2 integration.