This thesis addresses the open problem of automatically discovering hierarchical structure in reinforcement learning. Current algorithms for reinforcement learning fail to scale as problems become more complex. Many complex environments empirically exhibit hierarchy and can be modeled as interrelated subsystems, each in turn with hierarchic structure. Subsystems are often repetitive in time and space, meaning that they reoccur as components of different tasks or occur multiple times in different circumstances in the environment. A learning agent may sometimes scale to larger problems if it successfully exploits this repetition. Evidence suggests that a bottom up approach that repetitively finds building-blocks at one level of abstraction and uses them as background knowledge at the next level of abstraction, makes learning in many complex environments tractable. An algorithm, called HEXQ, is described that automatically decomposes and solves a multi-dimensional Markov decision problem (MDP) by constructing a multi-level hierarchy of interlinked subtasks without being given the model beforehand. The effectiveness and efficiency of the HEXQ decomposition depends largely on the choice of representation in terms of the variables, their temporal relationship and whether the problem exhibits a type of constrained stochasticity. The algorithm is first developed for stochastic shortest path problems and then extended to infinite horizon problems. The operation of the algorithm is demonstrated using a number of examples including a taxi domain, various navigation tasks, the Towers of Hanoi and a larger sporting problem. The main contributions of the thesis are the automation of (1)decomposition, (2) sub-goal identification, and (3) discovery of hierarchical structure for MDPs with states described by a number of variables or features. It points the way to further scaling opportunities that encompass approximations, partial observability, selective perception, relational representations and planning. The longer term research aim is to train rather than program intelligent agents

Reinforcement Learning (RL) is the study of learning agents that improve their performance from rewards and punishments. Most reinforcement learning methods optimize the discounted total reward received by an agent, while, in many domains, the natural criterion is to optimize the average reward per time step. In this thesis, we introduce a model-based average reward reinforcement learning method called "H-learning" and show that it performs better than other average reward and discounted RL methods in the domain of scheduling a simulated Automatic Guided Vehicle (AGV). We also introduce a version of H-learning which automatically explores the unexplored parts of the state space, while always choosing an apparently best action with respect to the current value function. We show that this "Auto-exploratory H-Learning" performs much better than the original H-learning under many previously studied exploration strategies. To scale H-learning to large state spaces, we extend it to learn action models and reward functions in the form of Bayesian networks, and approximate its value function using local linear regression. We show that both of these extensions are very effective in significantly reducing the space requirement of H-learning, and in making it converge much faster in the AGV scheduling task. Further, Auto-exploratory H-learning synergistically combines with Bayesian network model learning and value function approximation by local linear regression, yielding a highly effective average reward RL algorithm. We believe that the algorithms presented here have the potential to scale to large applications in the context of average reward optimization. / Graduation date:1996

Abstract not available

Reinforcement Learning (RL) is the study of agents that learn optimal behavior by interacting with and receiving rewards and punishments from an unknown environment. RL agents typically do this by learning value functions that assign a value to each state (situation) or to each state-action pair. Recently, there has been a growing interest in using hierarchical methods to cope with the complexity that arises due to the huge number of states found in most interesting real-world problems. Hierarchical methods seek to reduce this complexity by the use of temporal and state abstraction. Like most RL methods, most hierarchical RL methods optimize the discounted total reward that the agent receives. However, in many domains, the proper criteria to optimize is the average reward per time step. In this thesis, we adapt the concepts of hierarchical and recursive optimality, which are used to describe the kind of optimality achieved by hierarchical methods, to the average reward setting and show that they coincide under a condition called Result Distribution Invariance. We present two new model-based hierarchical RL methods, HH-learning and HAH-learning, that are intended to optimize the average reward. HH-learning is a hierarchical extension of the model-based, average-reward RL method, H-learning. Like H-learning, HH-learning requires exploration in order to learn correct domain models and optimal value function. HH-learning can be used with any exploration strategy whereas HAH-learning uses the principle of "optimism under uncertainty", which gives it a built-in "auto-exploratory" feature. We also give the hierarchical and auto-exploratory hierarchical versions of R-learning, a model-free average reward method, and a hierarchical version of ARTDP, a model-based discounted total reward method. We compare the performance of the "flat" and hierarchical methods in the task of scheduling an Automated Guided Vehicle (AGV) in a variety of settings. The results show that hierarchical methods can take advantage of temporal and state abstraction and converge in fewer steps than the flat methods. The exception is the hierarchical version of ARTDP. We give an explanation for this anomaly. Auto-exploratory hierarchical methods are faster than the hierarchical methods with ��-greedy exploration. Finally, hierarchical model-based methods are faster than hierarchical model-free methods. / Graduation date: 2003

Thesis (Ph. D.)--Rutgers University, 2009. / "Graduate Program in Computer Science." Includes bibliographical references (p. 100-104).

Not available / text

Not available / text

published_or_final_version / abstract / Electrical and Electronic Engineering / Doctoral / Doctor of Philosophy

Thesis (Ph. D.)--Rutgers University, 2009. / "Graduate Program in Computer Science." Includes bibliographical references (p. 238-261).

published_or_final_version / Industrial and Manufacturing Systems Engineering / Doctoral / Doctor of Philosophy