The family of Gradient Temporal-Difference (GTD) learning algorithms shares a promising property of being stable with both linear function approximation and off-policy training. The success of the GTD family requires a suitable set of features, which are unfortunately not always available in reality. To overcome this difficulty, regularization is often employed as an effective method for feature selection in reinforcement learning. In the present work, we propose and investigate a family of L1 regularized GTD learning algorithms.     «

The family of Gradient Temporal-Difference (GTD) learning algorithms shares a promising property of being stable with both linear function approximation and off-policy training. The success of the GTD family requires a suitable set of features, which are unfortunately not always available in reality. To overcome this difficulty, regularization is often employed as an effective method for feature selection in reinforcement learning. In the present work, we propose and investigate a family of L1...     »