Learning routing policies in a disrupted, congestible network with real-time information: An experimental approach

摘要

The roles of learning, inertia and real-time travel information on route choices in a highly disrupted network are investigated, based on data from a laboratory competitive route choice game. Routing policies instead of simple paths are treated as the subject of learning when real-time travel information is available, where a routing policy is defined as a contingency plan that maps realized traffic conditions to path choices. A learning model based on the power law of forgetting and reinforcement is applied to calculate the perceived travel times of alternative routing policies, based on which choices are made. A deterministic correction to the Logit choice model in a learning context is developed to account for overlapping routing policies.Model parameter estimates are obtained from maximizing the likelihood of making the observed choices on the current day based on choices from all previous days. Prediction performance is measured in terms of both one-step and full-trajectory predictions, based on observed choices up to today, in which one-step prediction entails predicting the next day's choice, while full-trajectory prediction entails predicting the next K days' choices. Three major conclusions are drawn. First, the routing policy learning model can capture travelers' learning and choice behavior better than a path-based model under real-time travel information, as it accounts for travelers' forward-looking capabilities. Secondly, inertia exists where travelers stick to previously chosen routes and do not necessarily minimize travel time. Inertia plays a dominant role in one-step prediction, and a less important role in full-trajectory prediction, suggesting that learning is more important in longer term prediction. Thirdly, relative importance of learning compared with inertia is more prominent in a less uncertain, but not close to deterministic, environment. Therefore, decreasing uncertainty by providing real-time information could encourage learning and potentially more optimal decisions for individuals and the system.