We demonstrate the use of cognitive architectures as a common framework for modeling AIs and human performers in a task. The mental model of each performer is revealed using cognitive salience, which is intended as a basis for explanation.
At it’s simplest you can use a memory model and instance-based learning in
pip install pyactup
The demo has full tutorials and examples.
This approach is intended for human-machine teaming where a human participant is invested in or participates in a task. The method adapts as the participant learns the task. It can be used to infer if an explanation is needed, the differences between AI and participant mental models, tailored to the individual.
We have applied this method in RL game context (gridworld) but more complex models could be developed for more complex environments.
The model is non-parametric (k-Nearest Neighbor) with terms to capture cognitive effects. It takes states (slot/value pairs) and maps to values (or actions).
The model is not “trained” but uses behavioral traces to supervise-learn behavior.
Normal k-Nearest Neighbor limitations.