We present a model-free reinforcement learning method for partially observable Markov decision problems. Our method estimates a likelihood gradient by sampling directly in parameter space, which leads to lower variance gradient estimates than obtained by regular policy gradient methods. We show that for several complex control tasks, including robust standing with a humanoid robot, this method outperforms well-known algorithms from the fields of standard policy gradients, finite difference methods and population based heuristics. We also show that the improvement is largest when the parameter samples are drawn symmetrically. Lastly we analyse the importance of the individual components of our method by incrementally incorporating them into the other algorithms, and measuring the gain in performance after each step.

Abstract. We present a model-free reinforcement learning method for partially observable Markov decision problems. Our method estimates a likelihood gradient by sampling directly in parameter space, which leads to lower variance gradient estimates than those obtained by policy gradient methods such as REINFORCE. For several complex control tasks, including robust standing with a humanoid robot, we show that our method outperforms well-known algorithms from the fields of policy gradients, finite difference methods and population based heuristics. We also provide a detailed analysis of the differences between our method and the other algorithms.

No abstract

No abstract

Abstract. In most real-world information processing problems, data is not a free resource; its acquisition is rather time-consuming and/or expensive. We investigate how these two factors can be included in supervised classification tasks by deriving classification as a sequential decision process and making it accessible to Reinforcement Learning. Our method performs a sequential feature selection that learns which features are most informative at each timestep, choosing the next feature depending on the already selected features and the internal belief of the classifier. Experiments on a handwritten digits classification task show significant reduction in required data for correct classification, while a medical diabetes prediction task illustrates variable feature cost minimization as a further property of our algorithm.