Experience-Oriented Artificial Intelligence

AI is at an impasse. It is stuck, or downsizing. Unable to build large, ambitious systems because no means to manage complexity. Now people manage complexity, but a large AI must do it itself. An AI must be able to tell for itself when it is right and when it is wrong. Experience is the route to this... Experience should be at the center of AI. It is what AI is about. It is the data of AI, yet it has been sidelined. An AI must be able to tell for itself when it is right and when it is wrong. Experience plays a central role in the problem of artificial intelligence. If intelligence is a computation, then the temporal stream of sensations is its input, and the temporal stream of actions is its output. These two intermingled time series are both the basis for all intelligent decision making and the basis for assessing it. Experience waits for neither man nor machine. Its events occur in an unalterable order and pace. Sensory signals may require quick action, or a more deliberate response. An action taken cannot be retracted. the temporal structure of experience may be the single most important computational feature of the problem of artificial intelligence. Nevertheless, experience has played a less than salient role in the field of artificial intelligence. Artificial intelligence has often dealt with subjects such as inference, diagnosis and problem-solving in such a way as to minimize the impact of real-time sensation and action. It is hard to discern any meaningful role for experience in classical question-answering AI systems. These systems may help people predict and control their experience, but the systems themselves have none. Robotics has always been an important exception, but even there experience and time play less of a role than might have been anticipated. Motor control is dominated by planning methods that emphasize trajectories and kinematics over dynamics. Computer vision research is concerned mostly with static images, or with open-loop streams of images with little role for action. Machine learning is dominated by methods which assume independent, identically distributed data—data in which order is irrelevant and there is no action. Recent trends in artificial intelligence can be seen as in part a shift in orientation towards experience. The “agent oriented” view of AI can be viewed in this light. Probabilistic models such as Markov decision processes, dynamic Bayes networks, and reinforcement learning are also part of the modern trend towards recognizing a primary role for temporal data and action. A natural place to begin exploring the role of experience in artificial intelligence is in knowledge representation. Knowledge is critical to the performance of successful AI systems, from the knowledgebase of a diagnosis system to the evaluation function of a chess-playing program to the map and sensor model of a navigating robot. Intelligence itself can be defined as the ability to maintain a very large body of knowledge and apply it effectively and flexibly to new problems.