An Adaptive Feedback System for Agent and Human Learning

点击翻译按钮获取中文摘要

Studies of technological innovation show that most new ideas need constructive feedback to become successful. Currently, there are no user-adaptive feedback systems that can be embedded in any learning and ideation environment. My research aims to discover fundamental principles for user-adaptive feedback systems that provide constructive feedback and to design and implement a system to validate these principles. My ultimate research goal is to create a general-purpose feedback system that can be embedded into a structured (with known rules) or unstructured (with rules to be discovered) environment and can learn to extract the rules of that environment. The proposed research program builds on my NSERC-funded doctoral program devising 1) a reinforcement learning (RL) algorithm, ALeRT, which enabled agents in games to learn adaptively, and 2) a model of collaborative agent behaviours. It also builds on my postdoctoral research, when I blended artificial intelligence with education to develop an intelligent rule-based feedback system. I discovered the principle that learners who seek negative feedback perform better on tasks and standardized tests, learning more than those seeking favourable feedback. My short-term goal is to identify such principles by creating a user-adaptive feedback system that generates constructive feedback for a structured domain-knowledge content. My medium-term goal is to create a system that generates feedback for an unstructured domain-knowledge content. My long-term goal is to enable multiple agents to collaboratively solve problems. Two PhDs, 1 MSc and 1 undergraduate per year will be trained throughout this program. This research includes modeling and experimental streams in a well-rounded methodology to create user-adaptable feedback agents. First, I will devise an RL algorithm enabling agents to increase their learning rate in a structured environment. I will discover a utility function to evaluate an agent's actions following feedback and the agent's performance. I will extend my ALeRT algorithm, so agents can learn to prioritize feedback. Second, I will infer the rules of an unstructured environment. Based on deep RL techniques, the value function of each action would then be used to populate the action space and the rules of the system. Third, I will extend the multiagent collaborative behaviour model I developed to enable agents to exchange feedback in a scalable way to achieve a common goal. I will validate my approach by embedding the feedback system in a different environment (e.g., the UofA massive open online course, Problem Solving, Programming, and Video Games). The project will deepen our understanding and guide research on user-adaptive systems in which agents learn to improve their decision-making that is crucial in developing autonomous learners. It will also contribute significantly to training highly-qualified personnel for successful and innovative academic or industry careers in Canada.