Representing Knowledge as Predictions (and State as Knowledge)

This paper shows how a single mechanism allows knowledge to be constructed layer by layer directly from an agent's raw sensorimotor stream. This mechanism, the General Value Function (GVF) or "forecast," captures high-level, abstract knowledge as a set of predictions about existing features and knowledge, based exclusively on the agent's low-level senses and actions. Thus, forecasts provide a representation for organizing raw sensorimotor data into useful abstractions over an unlimited number of layers--a long-sought goal of AI and cognitive science. The heart of this paper is a detailed thought experiment providing a concrete, step-by-step formal illustration of how an artificial agent can build true, useful, abstract knowledge from its raw sensorimotor experience alone. The knowledge is represented as a set of layered predictions (forecasts) about the agent's observed consequences of its actions. This illustration shows twelve separate layers: the lowest consisting of raw pixels, touch and force sensors, and a small number of actions; the higher layers increasing in abstraction, eventually resulting in rich knowledge about the agent's world, corresponding roughly to doorways, walls, rooms, and floor plans. I then argue that this general mechanism may allow the representation of a broad spectrum of everyday human knowledge.

T-Learning

Traditional Reinforcement Learning (RL) has focused on problems involving many states and few actions, such as simple grid worlds. Most real world problems, however, are of the opposite type, Involving Few relevant states and many actions. For example, to return home from a conference, humans identify only few subgoal states such as lobby, taxi, airport etc. Each valid behavior connecting two such states can be viewed as an action, and there are trillions of them. Assuming the subgoal identification problem is already solved, the quality of any RL method---in real-world settings---depends less on how well it scales with the number of states than on how well it scales with the number of actions. This is where our new method T-Learning excels, by evaluating the relatively few possible transits from one state to another in a policy-independent way, rather than a huge number of state-action pairs, or states in traditional policy-dependent ways. Illustrative experiments demonstrate that performance improvements of T-Learning over Q-learning can be arbitrarily large.