We introduce a new algorithm based on linear programming that approximates the differential value function of an average-cost Markov decision process via a linear combination of pre-selected basis functions. The algorithm carries out a form of cost shaping and minimizes a version of Bellman error. We establish an error bound that scales gracefully with the number of states without imposing the (strong) Lyapunov condition required by its counterpart in [6]. We propose a path-following method that automates selection of important algorithm parameters which represent counterparts to the "state-relevance weights" studied in [6].

Motor control depends on sensory feedback in multiple modalities with different latencies. In this paper we consider within the framework of reinforcement learning how different sensory modalities can be combined and selected for real-time, optimal movement control. We propose an actor-critic architecture with multiple modules, whose output are combined using a softmax function. We tested our architecture in a simulation of a sequential reaching task. Reaching was initially guided by visual feedback with a long latency. Our learning scheme allowed the agent to utilize the somatosensory feedback with shorter latency when the hand is near the experienced trajectory. In simulations with different latencies for visual and somatosensory feedback, we found that the agent depended more on feedback with shorter latency.

Successful application of reinforcement learning algorithms often involves considerable handcrafting of the necessary nonlinear features to reduce the complexity of the value functions and hence to promote convergence of the algorithm. In contrast, the human brain readily and autonomously finds the complex features when provided with sufficient training. Recent work in machine learning and neurophysiology has demonstrated the role of the basal ganglia and the frontal cortex in mammalian reinforcement learning. This paper develops and explores new reinforcement learning algorithms inspired by neurological evidence that provides potential new approaches to the feature construction problem. The algorithms are compared and evaluated on the Acrobot task.

In this setting, TD learning is often simpler and more data-efficient than other methods.

Psychologists call behavior intrinsically motivated when it is engaged in for its own sake rather than as a step toward solving a specific problem of clear practical value. But what we learn during intrinsically motivated behavior is essential for our development as competent autonomous entities ableto efficiently solve a wide range of practical problems as they arise. In this paper we present initial results from a computational study of intrinsically motivated reinforcement learning aimed at allowing artificial agentsto construct and extend hierarchies of reusable skills that are needed for competent autonomy.

Agents learning to act in a partially observable domain may need to overcome the problem of perceptual aliasing - i.e., different states that appear similar but require different responses. This problem is exacerbated whenthe agent's sensors are noisy, i.e., sensors may produce different observationsin the same state. We show that many well-known reinforcement learning methods designed to deal with perceptual aliasing, suchas Utile Suffix Memory, finite size history windows, eligibility traces, and memory bits, do not handle noisy sensors well. We suggest a new algorithm, Noisy Utile Suffix Memory (NUSM), based on USM, that uses a weighted classification of observed trajectories. We compare NUSM to the above methods and show it to be more robust to noise.

We introduce a new algorithm based on linear programming that approximates the differential value function of an average-cost Markov decision process via a linear combination of pre-selected basis functions. The algorithm carries out a form of cost shaping and minimizes a version of Bellman error. We establish an error bound that scales gracefully with the number of states without imposing the (strong) Lyapunov condition required by its counterpart in[6]. We propose a path-following method that automates selection of important algorithm parameters which represent counterparts tothe "state-relevance weights" studied in [6].

In this paper, we consider Markov Decision Processes (MDPs) with error states. Error states are those states entering which is undesirable or dangerous. We define the risk with respect to a policy as the probability of entering such a state when the policy is pursued. We consider the problem of finding good policies whose risk is smaller than some user-specified threshold, and formalize it as a constrained MDP with two criteria. The first criterion corresponds to the value function originally given. We will show that the risk can be formulated as a second criterion function based on a cumulative return, whose definition is independent of the original value function. We present a model free, heuristic reinforcement learning algorithm that aims at finding good deterministic policies. It is based on weighting the original value function and the risk. The weight parameter is adapted in order to find a feasible solution for the constrained problem that has a good performance with respect to the value function. The algorithm was successfully applied to the control of a feed tank with stochastic inflows that lies upstream of a distillation column. This control task was originally formulated as an optimal control problem with chance constraints, and it was solved under certain assumptions on the model to obtain an optimal solution. The power of our learning algorithm is that it can be used even when some of these restrictive assumptions are relaxed.

In this paper, we confront the problem of applying reinforcement learning to agents that perceive the environment through many sensors and that can perform parallel actions using many actuators as is the case in complex autonomous robots. We argue that reinforcement learning can only be successfully applied to this case if strong assumptions are made on the characteristics of the environment in which the learning is performed, so that the relevant sensor readings and motor commands can be readily identified. The introduction of such assumptions leads to strongly-biased learning systems that can eventually lose the generality of traditional reinforcement-learning algorithms. In this line, we observe that, in realistic situations, the reward received by the robot depends only on a reduced subset of all the executed actions and that only a reduced subset of the sensor inputs (possibly different in each situation and for each action) are relevant to predict the reward. We formalize this property in the so called 'categorizability assumption' and we present an algorithm that takes advantage of the categorizability of the environment, allowing a decrease in the learning time with respect to existing reinforcement-learning algorithms. Results of the application of the algorithm to a couple of simulated realistic-robotic problems (landmark-based navigation and the six-legged robot gait generation) are reported to validate our approach and to compare it to existing flat and generalization-based reinforcement-learning approaches.

Autonomous helicopter flight represents a challenging control problem, with complex, noisy, dynamics. In this paper, we describe a successful application of reinforcement learning to autonomous helicopter flight.