Infrastructure Optimisation via Deep Neural Search

In this work, we investigated how to optimise the deployment process of the infrastructure for urban mobility systems, to achieve more robust operation under dynamics, e.g. we want to figure out when and where to deploy or close the stations, so profit can be maximised for a given budget, which is often a function of time.

In particular, we developed a deep neural search approach, which uses a reinforcement learning controller to sample possible deployment plans, and runs them through simulation to evaluate their performance, e.g. in terms of service coverage and net revenue. This is then used as reward signals for the controller, which steers it to propose better strategies in future iterations.

We’ve shown empirically that our approach could discover strategies that achieves better service coverage and net revenue than the heuristic-based algorithms, under different optimisation goals, and also consistently offers better performance over time.

For more details please see the following paper:

Deployment Optimization for Shared e-Mobility Systems with Multi-agent Deep Neural Search.
Man Luo, Bowen Du, Konstantin Klemmer, Hongming Zhu, Hongkai Wen.
In IEEE Transactions on Intelligent Transportation Systems (T-ITS), 2021.