Download files
Abstract
Complex situations are very much context dependent, thus agents whether human or artificial need to attain an awareness based on their present situation. An essential part of that awareness is the accurate and effective perception and understanding of the set of knowledge, skills, and characteristics that are needed to allow an agent to perform a specific task with high performance, or what we would like to name, Competency Awareness. The development of this awareness is essential for any further development of an agent s capabilities. This can be assisted by identifying the limitations in the so far developed expertise and consequently engaging in training processes that add the necessary knowledge to overcome those limitations. However, current approaches of competency and situation awareness rely on manual, lengthy, subjective, and intrusive techniques, rendering those approaches as extremely troublesome and ineffective when it comes to developing computerized agents within complex scenarios. Within the context of computer Go, which is currently a grand challenge to Artificial Intelligence, these issues have led to substantial bottlenecks that need to be addressed in order to achieve further improvements. Thus, the underlying principle of this work is that of the development of an automated, objective and non-intrusive methodology of Competency Assessment of decision-makers, which will practically aid in the understanding and provision of effective guidance to the development of improved decision-making capabilities, specifically within computer agents.
In this study, we propose a framework whereby a computational environment is used to study and assess the competency of a decision maker. We use the game of GO to demonstrate this functionality in an environment in which hundreds of human-played GO games are analysed. In order to validate the proposed framework, a series of experiments on a wide range of problems have been conducted. These experiments automatically: (1) measure and monitor the competency of Human Go players, (2) reveal and monitor the dynamics of Neuro-Evolution, and (3) integrate strategic domain knowledge into evolutionary algorithms. The experimental results show that: (1) the proposed framework is effective in measuring and monitoring the strategic competencies of human Go players and evolved Go neuro-players, and (2) is effective in guiding the development of improved Go neuro-players when compared to traditional approaches that lacked the integration of a strategic competency measurement.