A Gaussian process approach for contextual bandit-based battery replacement

New York University Shanghai, New York University, 200124, Shanghai, China

^* Corresponding author: tz2321@nyu.edu

Abstract

The sharing economy has recently become a distinctive business model in China, offering advantages such as low prices and ease of use. Shared electric vehicles have also become an essential part of urban transportation. This research aims to assist electric vehicle providers in optimizing battery replacement schedules, with the objective of minimizing operating losses while meeting evening peak demand. Efficient resource allocation is crucial to remain competitive with the established public transport system. This study proposes a multi-armed bandit (MAB) approach to identify optimal periods for inspection and battery replacement in new city launches, even without prior knowledge of user behavior patterns. Modifications are made to the traditional MAB algorithm, incorporating the lower confidence bound (LCB), contextual features, kernelization, and the Gaussian Process (GP) to enhance the Upper Confidence Bound (UCB) MAB in solving this problem. Unlike deep learning techniques, the MAB model offers a lightweight, efficient, and easy-to-deploy solution that adapts to dynamic scenarios even with limited training data. Results indicate that this method performs stably in cumulative regret and in selecting the optimal choice within a short timeframe. Adaptable to seasonal and weekend fluctuations, this optimized approach shows potential for enhancing operational strategies not only in shared transportation but also across other sectors of the sharing economy.