Masters Thesis Defense: Vlad Chiriacescu

Abstract: Within cognitive science, computational modeling based on cognitive architectures has been an important approach to addressing questions of human cognition and learning. Modeling issues such as limited expressivity in representing knowledge and lack of appropriate selection of model structure represent a challenge for existing architectures. Furthermore, latest research shows that the concepts of long-term memory, motivation and working memory are critical cognitive aspects but a unifying cognitive paradigm integrating those concepts hasn’t been previously achieved. Derived from a synthesis of neuroscience, cognitive science, psychology, and education, the Unified Learning Model (ULM) provides this integration by merging a statistical learning mechanism with a general learning architecture.

Based on the ULM cognitive principles, this thesis presents a novel computational architecture called C-ULM that addresses the modeling issues outlined above and introduces a novel computational integration of long-term memory, motivation and working memory. C-ULM is implemented as a multi-agent simulation where the agent communication is grounded on the actions of teaching and learning. Both communication actions consist of two main phases: allocating working memory for teaching or learning and using the working memory content in order to update the agent's long-term memory.

From a cognitive perspective, C-ULM provides a test of the viability of the learning mechanisms proposed in the ULM. In addition, as showcased by C-ULM experiments, it offers insights that lead to a better understanding of the human learning mechanisms especially in the cases of long-term learning and problem solving where data from human subjects is generally not available.

From a multi-agent perspective, it advances the literature by providing the first multi-agent based simulation that incorporates long-term memory, working memory, motivation and the relationships among them into an effective modeling framework. Furthermore, it offers the foundation for novel agent reasoning models and insights into modeling agent-to-agent knowledge transfer based on the principles of human learning and teaching processes.

Committee Members: Dr. LeenKiat Soh (Advisor), Dr. Ashok Samal, and Dr. Duane Schell