Combining self-motivation with planning and inference in a self-motivated cognitive agent framework

Thesis (Ph. D.)--University of Rochester. Dept. of Computer Science, 2013.Most work on self-motivated agents in Artificial Intelligence has focused on acquiring utility-optimizing mappings from states to actions. However, such mappings do not allow for explicit, reasoned anticipation and planned achievement of future states and pay-offs, based on symbolic knowledge about the environment and about the consequences of the agent's own behavior. In essence, such agents can behave only reflexively, not reflectively. Conversely, planning and reasoning have been viewed as aimed at fulfillment of explicitly specified user goals, without regard for the long-range utility of the planner or reasoner's choices. Moreover, most work assumes either a completely known environment or a predictable environment altered only by the agent's actions. We strive to integrate self-motivation into logical planning and reasoning, in a general self-motivated cognitive agent framework. Our self-motivated cognitive agents are capable of thinking ahead and planning flexibly in an incompletely known world, question-answering dialogues, mental modeling, introspection and reasoning over a substantial knowledge base, all while in unceasing pursuit of optimizing their own cumulative utility. While such agents employ reasoned exploration of feasible sequences of actions and corresponding states, they also behave opportunistically, seizing opportunities and recovering from failures, thanks to their continual plan updates and quest for rewards. Our framework allows for both specific and general (quantified) knowledge, and for epistemic predicates such as knowing--that and knowing--whether. Since realistic agents often possess incomplete knowledge of their world, the reasoning of our agents uses a prudently restricted version of the closed world assumption; this has consequences for epistemic reasoning, in particular positive and negative introspection. The planning operators allow for quantitative, gradual change and side effects such as the passage of time, changes in distances and rewards, and language production, using a uniform procedural attachment technique. Question-answering (involving introspection and reasoning) and experimental runs are shown for our particular agent ME in a simulated world, demonstrating the benefits of self-awareness as well as continual, deliberate and reward-driven planning. Our agents can also be easily configured as single or multiple goal-seeking agents, and as such perform quite favorably compared with recent experimental agents