Although reinforcement learning (RL) defines the learning problem in a more general setting than does supervised learning, making it a better fit to a broader spectrum of real-life learning tasks, it has been far less successful in gaining attention from other scientific disciplines and commercial ventures beyond machine learning. In an effort to help coax reinforcement learning out of the labs, we introduce a novel benchmark methodology and problem repository. Our approach involves collection and distribution of static training/testing datasets, similar to efforts in other disciplines such as the UCI repository for supervised learning [4]. The new framework allows for a fresh set of benchmarks grounded in measured data that are challenging, consistent, and realistic. We believe that the new approach can succeed as a focal point for empirical developments in the field. The current situation in reinforcement learning is akin to the pre-UCI era in supervised learning. New algorithms are validated with one or two artificial benchmark problems and often require problem-specific tweaking. Recently however, some impressive attempts have been made to grow reinforcement learning out of this model. The RL repository at the University of Massachusetts 1, the University of Alberta�s RL-Glue project 2 and the RL Competition are examples that have helped unify the evaluation of RL algorithms. However, the nature of the evaluation schemes of these methods have made it difficult to include actual measured data from real-life domains, like those found in the supervised-learning benchmarks. Creating an RL benchmark is not straightforward. For example, it is not apparent what data format should be used. In supervised learning, data can easily be provided as text files, usually in the form of a mapping from input space X to an output space Y. However, RL environ-

Oh la la

• Read
• Review
• Email
• Add to collection
• Download

Review

Your message has been sent.