Cold-Start Reinforcement Learning with Softmax Policy Gradient
This work addresses efficiency and simplicity challenges in reinforcement learning for structured output prediction, though it appears incremental as it builds on existing policy-gradient and maximum-likelihood approaches.
The paper tackles the issues of warm-start training and sample variance reduction in policy-gradient reinforcement learning by introducing a softmax value function method that eliminates both procedures, achieving competitive performance on automatic summarization and image captioning tasks.
Policy-gradient approaches to reinforcement learning have two common and undesirable overhead procedures, namely warm-start training and sample variance reduction. In this paper, we describe a reinforcement learning method based on a softmax value function that requires neither of these procedures. Our method combines the advantages of policy-gradient methods with the efficiency and simplicity of maximum-likelihood approaches. We apply this new cold-start reinforcement learning method in training sequence generation models for structured output prediction problems. Empirical evidence validates this method on automatic summarization and image captioning tasks.