Online Decision Transformer

Generative pretraining for sequence modeling has emerged as a unifying paradigm for machine learning in a number of domains and modalities, notably in language and vision (Radford et al., 2018; Chen et al., 2020; Brown et al., 2020; Lu et al., 2022). Recently, such a pretraining paradigm has been extended to offline reinforcement learning (RL) (Chen et al., 2021; Janner et al., 2021), wherein an agent is trained to autoregressively maximize the likelihood of trajectories in the offline dataset. During training, this paradigm essentially converts offline RL to a supervised learning problem (Schmidhuber, 2019; Srivastava et al., 2019; Emmons et al., 2021). However, these works present an incomplete picture as policies learned via offline RL are limited by the quality of the training dataset and need to be finetuned to the task of interest via online interactions. It remains an open question whether such supervised learning paradigm can be extended to online settings. Unlike language and perception, online finetuning for RL is fundamentally different from the pretraining phase as it involves data acquisition via exploration. The need for exploration renders traditional supervised learning objectives (e.g., mean squared error) for offline RL insufficient in the online setting. Moreover, it has been observed that for standard online algorithms, access to offline data can often have zero or even negative effect on the online performance (Nair et al., 2020). Hence, the overall pipeline for offline pretraining followed by online finetuning for RL policies needs a careful consideration of training objectives and protocols.