REPOSITORY > RESULTS

Abstract

Skill training in real-life circumstances for specific personnel (such as military, police, surgeons, and pilots) can be expensive, dangerous to the trainees and others, and time-consuming. Moreover, it is often impossible for the trainee to explore all the possible circumstances or deal with problems that might occur. In addition, the emotional and psychological states and well-being of people should be taken into consideration. Due to these reasons, extensive skill training using simulations for improving decision making has become common practice in a multitude of domains and professions.
In this dissertation, we present a novel training approach called PATDS (Performance Analysis of Training through Data-Farming Simulations) that introduces a training cycle, where each iteration identifies the weaknesses of the trainee and another round of training is performed to improve the performance on the identified problematic situations. This is achieved by constructing a simulated agent as a virtual copy of a human trainee and testing it in all circumstances to provide feedback and decision support for another round of human learning. The main concepts of the procedure resemble the boosting methodology from the machine learning domain, but instead of improving a learning model, the human surrogate and the human trainee are being improved.
PATDS is composed of three main phases: (i) behavior analysis; (ii) pattern selection and behavior cloning; and (iii) performance evaluation and identification of weaknesses. In the behavior analysis phase, the relevant behavior patterns, demonstrated by the trainee in simulations, are identified, extracted, and described. These patterns are obtained using the HMASDA algorithm, which creates a set of identified patterns from basic observations. In the second phase, all obtained patterns are clustered in order to obtain a smaller subset of key patterns describing the physical and cognitive behavior of observed agents. Moreover, the patterns are transformed into one or many models of surrogate agents of the trainees, with the goal of enabling as accurate simulated behavior of the trainees as possible in a reasonably short learning period. In this way, artificial copies of the human trainees are obtained, which are capable of active performance under general conditions. In the final phase, the created agents are evaluated within data-farming experiments on a variety of simulation parameters to discover the key weaknesses in the trainees' learned strategies.
The motivation behind PATDS is to replace the need for the trainees to extensively test a huge set of possible specific conditions and scenarios, including various behaviors of the competing agents. The automated process can, in turn, test all options in a reasonable amount of time and detect circumstances causing weak performances. Consequently, in each subsequent round of training, the trainee can improve the behavior based on the identified weaknesses of the surrogate copy of the trainee. We evaluated the proposed approach on two asymmetric urban-terrain threat scenarios, where a squad of security forces interacts with hostile crowds under various conditions. We first show that the discovered behavior patterns are rich enough to successfully reproduce the observed behavior of the trainee, and then demonstrate how the data-farming-based evaluation on the complete simulation parameter space is beneficial for improving the trainee’s capabilities.