Task-Agnostic Online Reinforcement Learning with an Infinite Mixture of Gaussian Processes

Handle unknown tasks:

1. Meta-learning achieves quick adaptation and generalization with learned inductive bias. But it assumes that the tasks for training and testing are independently sampled from the same accessible distribution.
2. Continual learning aims to solve a sequence of tasks with clear task delineations while avoiding catastrophic forgetting.

In real-world, the assumptions of both methods are not met,

• The mutual knowledge transfer in meta-learning may degrade the generalization performance.
• Task distribution modeling those interactions are complex to determine.

This work is to solve nonstationary online problems where the task boundaries and the number of tasks are unknown by proposing a model-based reinforcement learning (RL) method that does not require a pre-trained model.

• either create a new model for unseen dynamics or recalls an old model for encountered dynamics.
• task recognition => update model parameter via conjugate gradient

###

---

• ← Previous Post
• Next Post →