Using reinforcement learning to minimize taxi idle times

Abstract

Taxis spend a large amount of time idle, searching for passengers. The routes vacant taxis should follow in order to minimize their idle times are hard to calculate; they depend on complex effects like passenger demand, traffic conditions, and inter-taxi competition. Here we explore if reinforcement learning (RL) can be used for this purpose. Using real-world data from three major cities, we show RL-taxis can indeed learn to minimize their idle times in different environments. In particular, a single RL-taxi competing with a population of regular taxis learns to out-perform its rivals.

Keywords:

• Data science
• machine learning
• networks
• optimization
• reinforcement learning
• taxis

Disclosure statement

No potential conflict of interest was reported by the author(s).

 

Figure A2. Subset of San Francisco street network.

Figure A3. Subset of Singapore street network.

Figure A4. Trip generation probabilities p_i from each city. The grid-world p_i are chosen uniformly at random.

Notes

1 We have abused notation slightly here by using $\pi (j|i)$ to denote the probability of moving to node j when at node i. This is abusive since we earlier defined the policy is term of a state s and action a, $\pi (s,a),$ so by $\pi (j|i)$ we mean $\pi (s=i,a\prime )$ where $a\prime$ denotes to node j is selected when at node i.