Constraint-Aware Meta-Learning, with Applications to Traffic Flow Prediction

Filardi, Vincent

Etd

Constraint-Aware Meta-Learning, with Applications to Traffic Flow Prediction

Public

We consider the general problem of training a machine learning model via a meta-learning approach, on a set of tasks T1, . . . , Tn where the data D1, . . . , Dn in each task satisfies a different set of inequality constraints C1, . . . , Cn. For each task Ti, we are given a dataset Di and a set of inequality constraints Ci, which the data is known to satisfy. The goal is to train the meta- learning model in such a way that it has a high prediction accuracy on each task without over-fitting the data. In the traditional meta-learning framework, task-specific constraint information Ci is not taken into account, and the learner must adapt the model to each task Ti using only the data Di. In contrast, we propose a “constraint-aware” meta-learning framework, where we empower the meta-learner and adaptive learner to take into account both the data Di as well as the constraints Ci which the data satisfies for any given task. This can potentially allow the trained model to achieve a much higher accuracy without over-fitting on tasks Ti where the dataset Di is very small (or where there is no task-specific data available for the task Ti), provided the task-specific constraints Ci are known. We apply our general framework to the problem of predicting the traffic flow of vehicles on different road networks. Changes to a road network present challenges to urban space and its mobility patterns. Here, the task Ti corresponds to the problem of predicting traffic flow on a given road network Gi, and the topology Gi of the road network imposes constraints Ci on the vehicle drivers. Our model learns how drivers respond to the constraints Ci imposed by the topology of the road network, by comparing traffic flow data from different road networks with different topologies Gi. We apply our model to a synthetic dataset generated by a traffic model with drivers that use shortest-path decision rules. When applied to this dataset, we observe that our model achieves a lower prediction error when compared to baseline models which do not take into account the constraints on the given tasks.

Creator