Knowledge Representation, Learning, and Problem Solving for General Intelligence

Abstract

For an intelligent agent to be fully autonomous and adaptive, all aspects of intelligent processing from perception to action must be engaged and integrated. To make the research tractable, a good approach is to address these issues in a simplified micro-environment that nevertheless engages all the issues from perception to action. We describe a domain independent and scalable representational scheme and a computational process encoded in a computer program called LEPS (Learning from Experience and Problem Solving) that addresses the entire process of learning from the visual world to the use of the learned knowledge for problem solving and action plan generation. The representational scheme is temporally explicit and is able to capture the causal processes in the visual world naturally and directly, providing a unified framework for unsupervised learning, rule encoding, problem solving, and action plan generation. This representational scheme allows concepts to be grounded in micro-activities (elemental changes in space and time of the features of objects and processes) and yet allow scalability to more complex activities like those encountered in the real world.