Learning belief connections in a model for situation awareness

Abstract

Situational awareness is critical in many human tasks, especially in cases where humans have to make decisions fast and where the result of their decisions might affect their life. This paper addresses the problem of learning optimal values for the parameters of a situational awareness model. The model is a complex network with nodes connected by links with weights, which connect observations to simple beliefs, such as "there is a contact", to complex belief, such as "the contact is hostile", and to future beliefs, such as "it is possible the pilot is being targeted". The model has been built and validated by human experts in the domain of F16 fighter pilots and is used to study human decision making. Given the complexity of the model, there is a need to learn appropriate weights for the connections, which, in turn, affect the activation levels of the beliefs. We propose the use of a genetic algorithm and of a sensitivity based approach to learn the weights in the model. Extensive experimental results are included. © 2011 Springer-Verlag.