As the development of autonomous vehicle technologies, autonomous taxis have been regarded as an emerging traffic mode. Although promising, autonomous taxis suffer inefficiency when searching for passengers due to the mismatch between fluctuated passenger demand and empty taxis. Since few research studies how an autonomous taxi search for passengers, this paper proposes two passenger searching algorithms by means of self-learning approaches: Bayesian-learning algorithm and L-Drive (Landmark-Drive) algorithm. Results reveal that the Bayesian-learning algorithm comes with hysteresis when the passenger demand is dynamically changing, while the L-Drive algorithm performs well when passenger demand changes suddenly. The reason for their difference is that the Bayesian-learning algorithm only hopes to shorten searching time to receive passengers, ignoring the spatial distribution of passenger demand. The proposed algorithms enable taxi agents to make decisions to find passengers quickly. They can significantly reduce deadhead mileage and passenger waiting time, as well as improve passenger service rate.

CICTP