Sequence Tutor: Conservative Fine-Tuning of Sequence Generation Models with KL-control
This addresses the challenge of generating high-quality, structured sequences in domains like music and molecular design, though it is incremental as it builds on existing RL and KL-control techniques.
The paper tackles the problem of improving sequence generation quality in RNNs while preserving learned data patterns and diversity, by proposing a method that combines pre-training with reinforcement learning and KL-control, resulting in enhanced properties for musical melodies and molecular generation.
This paper proposes a general method for improving the structure and quality of sequences generated by a recurrent neural network (RNN), while maintaining information originally learned from data, as well as sample diversity. An RNN is first pre-trained on data using maximum likelihood estimation (MLE), and the probability distribution over the next token in the sequence learned by this model is treated as a prior policy. Another RNN is then trained using reinforcement learning (RL) to generate higher-quality outputs that account for domain-specific incentives while retaining proximity to the prior policy of the MLE RNN. To formalize this objective, we derive novel off-policy RL methods for RNNs from KL-control. The effectiveness of the approach is demonstrated on two applications; 1) generating novel musical melodies, and 2) computational molecular generation. For both problems, we show that the proposed method improves the desired properties and structure of the generated sequences, while maintaining information learned from data.