Browsing by Subject "reinforcement learning"
Now showing 1 - 3 of 3
Results Per Page
Sort Options
Item Discretization and Approximation Methods for Reinforcement Learning of Highly Reconfigurable Systems(2010-07-14) Lampton, Amanda K.There are a number of techniques that are used to solve reinforcement learning problems, but very few that have been developed for and tested on highly reconfigurable systems cast as reinforcement learning problems. Reconfigurable systems refers to a vehicle (air, ground, or water) or collection of vehicles that can change its geometrical features, i.e. shape or formation, to perform tasks that the vehicle could not otherwise accomplish. These systems tend to be optimized for several operating conditions, and then controllers are designed to reconfigure the system from one operating condition to another. Q-learning, an unsupervised episodic learning technique that solves the reinforcement learning problem, is an attractive control methodology for reconfigurable systems. It has been successfully applied to a myriad of control problems, and there are a number of variations that were developed to avoid or alleviate some limitations in earlier version of this approach. This dissertation describes the development of three modular enhancements to the Q-learning algorithm that solve some of the unique problems that arise when working with this class of systems, such as the complex interaction of reconfigurable parameters and computationally intensive models of the systems. A multi-resolution state-space discretization method is developed that adaptively rediscretizes the state-space by progressively finer grids around one or more distinct Regions Of Interest within the state or learning space. A genetic algorithm that autonomously selects the basis functions to be used in the approximation of the action-value function is applied periodically throughout the learning process. Policy comparison is added to monitor the state of the policy encoded in the action-value function to prevent unnecessary episodes at each level of discretization. This approach is validated on several problems including an inverted pendulum, reconfigurable airfoil, and reconfigurable wing. Results show that the multi-resolution state-space discretization method reduces the number of state-action pairs, often by an order of magnitude, required to achieve a specific goal and the policy comparison prevents unnecessary episodes once the policy has converged to a usable policy. Results also show that the genetic algorithm is a promising candidate for the selection of basis functions for function approximation of the action-value function.Item Scaling reinforcement learning to the unconstrained multi-agent domain(2009-06-02) Palmer, VictorReinforcement learning is a machine learning technique designed to mimic the way animals learn by receiving rewards and punishment. It is designed to train intelligent agents when very little is known about the agent?s environment, and consequently the agent?s designer is unable to hand-craft an appropriate policy. Using reinforcement learning, the agent?s designer can merely give reward to the agent when it does something right, and the algorithm will craft an appropriate policy automatically. In many situations it is desirable to use this technique to train systems of agents (for example, to train robots to play RoboCup soccer in a coordinated fashion). Unfortunately, several significant computational issues occur when using this technique to train systems of agents. This dissertation introduces a suite of techniques that overcome many of these difficulties in various common situations. First, we show how multi-agent reinforcement learning can be made more tractable by forming coalitions out of the agents, and training each coalition separately. Coalitions are formed by using information-theoretic techniques, and we find that by using a coalition-based approach, the computational complexity of reinforcement-learning can be made linear in the total system agent count. Next we look at ways to integrate domain knowledge into the reinforcement learning process, and how this can signifi-cantly improve the policy quality in multi-agent situations. Specifically, we find that integrating domain knowledge into a reinforcement learning process can overcome training data deficiencies and allow the learner to converge to acceptable solutions when lack of training data would have prevented such convergence without domain knowledge. We then show how to train policies over continuous action spaces, which can reduce problem complexity for domains that require continuous action spaces (analog controllers) by eliminating the need to finely discretize the action space. Finally, we look at ways to perform reinforcement learning on modern GPUs and show how by doing this we can tackle significantly larger problems. We find that by offloading some of the RL computation to the GPU, we can achieve almost a 4.5 speedup factor in the total training process.Item Scaling Up Reinforcement Learning without Sacrificing Optimality by Constraining Exploration(2012-12-05) Mann, Timothy 1984-The purpose of this dissertation is to understand how algorithms can efficiently learn to solve new tasks based on previous experience, instead of being explicitly programmed with a solution for each task that we want it to solve. Here a task is a series of decisions, such as a robot vacuum deciding which room to clean next or an intelligent car deciding to stop at a traffic light. In such a case, state-of-the-art learning algorithms are difficult to employ in practice because they often make thou- sands of mistakes before reliably solving a task. However, humans learn solutions to novel tasks, often making fewer mistakes, which suggests that efficient learning algorithms may exist. One advantage that humans have over state- of-the-art learning algorithms is that, while learning a new task, humans can apply knowledge gained from previously solved tasks. The central hypothesis investigated by this dissertation is that learning algorithms can solve new tasks more efficiently when they take into consideration knowledge learned from solving previous tasks. Al- though this hypothesis may appear to be obviously true, what knowledge to use and how to apply that knowledge to new tasks is a challenging, open research problem. I investigate this hypothesis in three ways. First, I developed a new learning algorithm that is able to use prior knowledge to constrain the exploration space. Second, I extended a powerful theoretical framework in machine learning, called Probably Approximately Correct, so that I can formally compare the efficiency of algorithms that solve only a single task to algorithms that consider knowledge from previously solved tasks. With this framework, I found sufficient conditions for using knowledge from previous tasks to improve efficiency of learning to solve new tasks and also identified conditions where transferring knowledge may impede learning. I present situations where transfer learning can be used to intelligently constrain the exploration space so that optimality loss can be minimized. Finally, I tested the efficiency of my algorithms in various experimental domains. These theoretical and empirical results provide support for my central hypothesis. The theory and experiments of this dissertation provide a deeper understanding of what makes a learning algorithm efficient so that it can be widely used in practice. Finally, these results also contribute the general goal of creating autonomous machines that can be reliably employed to solve complex tasks.