|Abstract:||For many tasks such as text categorization and control of robotic systems, state-of-the art learning systems can produce results comparable in accuracy to those of human subjects. However, the amount of training data needed for such systems can be prohibitively large for many practical problems. A text categorization system, for example, may need to see many text postings manually tagged with their subjects before it learns to predict the subject of the next posting with high accuracy. A reinforcement learning (RL) system learning how to drive a car needs a lot of experimentation with the actual car before acquiring the optimal policy. An optimizing compiler targeting a certain platform has to construct, compile, and execute many versions of the same code with different optimization parameters to determine which optimizations work best. Such extensive sampling can be time-consuming, expensive (in terms of both expense of the human expertise needed to label data and wear and tear on the robotic equipment used for exploration in case of RL), and sometimes dangerous (e.g., an RL agent driving the car off the cliff to see if it survives the crash). The goal of this work is to reduce the amount of training data an agent needs in order to learn how to perform a task successfully. This is done by providing the system with prior knowledge about its domain. The knowledge is used to bias the agent towards useful solutions and limit the amount of training needed.
We explore this task in three contexts: classification (determining the subject of a newsgroup posting), control (learning to perform tasks such as driving a car up the mountain in simulation), and optimization (optimizing performance of linear algebra operations on different hardware platforms). For the text categorization problem, we introduce a novel algorithm which efficiently integrates prior knowledge into large margin classification. We show that prior knowledge simplifies the problem by reducing the size of the hypothesis space. We also provide formal convergence guarantees for our algorithm. For reinforcement learning, we introduce a novel framework for defining planning problems in terms of qualitative statements about the world (e.g., ``the faster the car is going, the more likely it is to reach the top of the mountain''). We present an algorithm based on policy iteration for solving such qualitative problems and prove its convergence. We also present an alternative framework which allows the user to specify prior knowledge quantitatively in form of a Markov Decision Process (MDP). This prior is used to focus exploration on those regions of the world in which the optimal policy is most sensitive to perturbations in transition probabilities and rewards. Finally, in the compiler optimization problem, the prior is based on an analytic model which determines good optimization parameters for a given platform. This model defines a Bayesian prior which, combined with empirical samples (obtained by measuring the performance of optimized code segments), determines the maximum-a-posteriori estimate of the optimization parameters.