An integrated neural-symbolic cognitive agent architecture for training and assessment in simulators

Training and assessment of complex tasks has always been a complex task in itself. Training simulators can be used for training and assessment of low-order skills. High-order skills (e.g. safe driving, leadership, tactical manoeuvring, etc.) are generally trained and assessed by human experts, due to its complex nature (i.e. many temporal relations, biased behaviour and poorly documented). This paper proposes a new cognitive agent architecture that is able to model this complex behaviour and use it for the assessment and training of both low- and high-order skills. Therefore the agent integrates learning from observation, temporal logic and probalistic reasoning in a unified architecture that is based on Neural-Symbolic Learning and Reasoning. This so-called Neural Symbolic Cognitive Agent (NSCA) architecture combines encoding temporal logic based expert knowledge and learning new knowledge by observing experts and trainees during task execution in a simulator. The learned knowledge can be extracted in temporal logic rules for validation. Learning and reasoning is done using a Recurrent Temporal Restricted Boltzmann Machine (RTRBM). For training organizations, this provides a quicker, cost-saving and more objective evaluation of the trainee in simulation-based training. A prototype NSCA has been developed and tested as part of a three-year research project on assessment in driving simulators for training and certification, and will be tested in various other domains, such as jetfighter pilot training and strategic command and control training.