Vinsamlegast notið þetta auðkenni þegar þið vitnið til verksins eða tengið í það: http://hdl.handle.net/1946/39956
Artificial Intelligence systems that need to operate in the physical world require a firm understanding of causal relationships to efficiently carry out their tasks. At present, few cognitive architectures, artificial intelligence, or control systems consider causal relationships between phenomena for achieving goals in the real world, and few if any tests exist to verify understanding of these relationships in intelligent learners.
The goal of this thesis is to outline a theory of tasks grounded in models of real-time causal relationships, intended to be used to analyze tasks of varying complexity and build tests for evaluating the ability of artificial systems to learn, perform, and understand complex real-world task-environments. Like an engineer can design a bridge with certain physical properties without actually building it, such a task theory, grounded in physics, should enable researchers to evaluate and compare tasks and systems without having to resort physical experiments.
Historically, research on the principles of tasks has been scarce, and causal relationships have never been a primary focus in those attempts. Few tests of causal understanding have been proposed thus far and the lack of a proper theory of tasks limits their utility. No comprehensive theory of tasks exists. We envision that such a theory would enable comparison of similar and different asks, calculation of task difficulty for particular learners, and prescriptive ways for modifying existing tasks to make them more tractable for particular performers and environments. Comparison of tasks would be invaluable when evaluating and considering the pros and cons of various approaches to AI systems and learners.
We base our theory of tasks on previous work on causal analysis and cumulative learning, proposing a method for computing complexity and comparability using causal interpretation of directed acyclic graphs (DAGs). The theory allows tasks to be analyzed along different dimensions and can be composed and decomposed into subgoals. It is intended to be used for evaluating and testing intelligent agents in various ways.