Continuous Adaptation with Online Meta-Learning for Non-Stationary Target Regression Tasks

Taku Yamagata,Raúl Santos-Rodríguez,Peter Flach

doi:10.3390/signals3010006

Abstract

Most environments change over time. Being able to adapt to such non-stationary environments is vital for real-world applications of many machine learning algorithms. In this work, we propose CORAL, a computationally efficient regression algorithm capable of adapting to a non-stationary target. CORAL is based on Bayesian linear regression with a sliding window and offline/online meta-learning. The sliding window makes our model focus on the recently received data and ignores older observations. The meta-learning approach allows us to learn the prior distribution of the model parameters. It speeds up the model adaptation, complements the sliding window’s drawback, and enhances the performance. We evaluate CORAL on two tasks: a toy problem and a more complex blood glucose level prediction task. Our approach improves the prediction accuracy for the non-stationary target significantly while also performing well for the stationary target. We show that the two components of our method work in a complementary fashion to achieve this.

Highlights

Adapting to a non-stationary environment is important for applying many machine learning algorithms to the real world [1]
We review the offline meta-learning algorithm proposed as a part of the Adaptive Learning for Probabilistic Connectionist Architecture (ALPaCA) algorithm [13]
The results suggest that the offline meta-learning improves a lot in the root mean squared error (RMSE)

Summary

Introduction

Adapting to a non-stationary environment is important for applying many machine learning algorithms to the real world [1]. We consider a task performed on a stream of data in such a non-stationary environment. This requires the model to learn as it receives data to perform better for the new incoming data. We assume that there is some prior knowledge (or data) about the task. The knowledge or data do not need to be accurate and precise since our model exploits them as long as they give an indication about the task. The data may be synthetically generated with a simulator, which is unlikely to match the real target. Having prior knowledge (or data) is not a mandatory requirement for using our method—we can incorporate prior knowledge whenever it exists

Methods

Results

Conclusion