Sequence Tutor: Conservative Fine-Tuning of Sequence Generation Models with KL-control

TL;DR

This paper introduces Sequence Tutor, a method combining pre-trained RNNs with reinforcement learning via KL-control to enhance sequence quality and structure while preserving learned information, demonstrated on music and molecular generation.

Contribution

It presents a novel off-policy RL approach using KL-control for fine-tuning RNNs, improving sequence quality without losing original data information.

Findings

Improved sequence structure and quality in music and molecular generation

Maintained diversity and learned information from data

Demonstrated effectiveness on two distinct applications

Abstract

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.