Joint pavement design & maintenance with pre-trained RL
One deep RL agent that designs and maintains pavements across many projects, with MIT CSHub.
Pavement design and maintenance are usually decided separately: engineers pick a structure, and an agency later decides when and how to repair it. This project optimizes both together.
In collaboration with the MIT Concrete Sustainability Hub (Dr. Randolph Kirchain and Dr. Haoran Li), we:
- formulated CSHub’s life-cycle cost and environmental models as a sequential decision problem;
- trained a dueling double deep Q-network (D3QN) with prioritized experience replay on randomized pavement contexts (structure, materials, traffic, stochastic roughness deterioration), so a single maintenance and rehabilitation policy serves many projects without retraining;
- reused the trained value function as a zero-shot surrogate to screen thousands of candidate concrete pavement designs, then selected the final design with a Monte Carlo elimination tournament, all on a single workstation;
- extended the agent to take the discount rate as an input, so agencies using different rates can apply it directly.
The life-cycle cost scope covers construction, maintenance, pavement–vehicle interaction, and end of life, using neural-network surrogates of mechanistic-empirical distress models.
This work has been accepted for presentation at the Transportation Research Board 106th Annual Meeting (January 2027).