Learning Riders Preferences In Ridesharing Platforms

"Ridesharing platforms allow people to commute more efficiently. Ridesharing can be beneficial since it can reduce the travel expenses for individuals as well as decrease the overall traffic gridlocks. One of the key aspects of ridesharing platforms is for riders to find suitable partners to share the ride. Thus, the riders need to be matched to other riders/drivers. From the social perspective, a rider may prefer to share the ride with certain individuals as opposed to other riders. This leads to the rider having preferences over the other riders. A matching based on social welfare indicates the quality of the rides. Our goal is to maximize social welfare or the quality of rides for all riders. In order to match the riders, we need to know the preferences of the riders. However, the preferences are often unknown. To tackle these situations, we introduce a ridesharing model that implements reinforcement learning algorithms to learn the utilities of the riders based on the riders' previous experiences. We investigate a variety of measures for assessing social welfare, including utilitarian, egalitarian, Nash, and leximin social welfare. Additionally, we also compute the number of strong and weak blocking pairs in each socially optimal matching to compare the stability of these matchings. We provide a comparison between two reinforcement learning algorithms: [epsilon]-greedy and UCB1, for learning utilities of the riders, maximizing social welfare, and the number of blocking pairs in the socially optimal matching. The [epsilon]-greedy algorithm with [epsilon]=0.1 provides the maximum accuracy in learning the utilities of the riders as compared to [epsilon]=0.0, [epsilon]=0.01, and UCB1 algorithm. It also provides a fewer number of blocking pairs suggesting more stability in the socially optimal matching than other reinforcement learning algorithms. However, the UCB1 algorithm outperforms all other reinforcement learning algorithms to provide maximum welfare in socially optimal matchings."--Abstract.