Structural and Functional Representativity of GANs for Data Generation in Sequential Decision Making

摘要

In many sequential decision making problems progress is predominantly based on artificial data sets. This can be attributed to insufficient access to real data. Here we propose to mitigate this by using generative adversarial networks (GANs) to generate representative data sets from real data. Specifically, we investigate how GANs can generate training data for reinforcement learning (RL) problems. We distinguish structural properties (does the generated data follow the distribution of the original data), functional properties (is there a difference between the evaluation of policies for generated and real life data), and show that with a relatively small number of data points (a few thousand) we can train GANs that generate representative data for classical control RL environments.