City Science secured funding from Innovate UK to advance Phase 2 of our AI for Mass Model Automation with the University of Exeter. In Phase 1, we demonstrated the feasibility of automating travel demand matrices for transport modelling. In Phase 2, we continued to develop AI-driven optimisation techniques to reduce time and effort and increase accuracy in transport modeling processes, specifically mode choice model calibration.
Overview
Transport modeling is a highly time-consuming and resource-intensive process. We identified variable demand modeling as an area of significant time and resource investment and a potential gap in the model automation market. Existing methods are manual, involving trial and error and often fail to identify the best solutions.
In collaboration with the University of Exeter, City Science integrated AI optimisation techniques for variable demand modelling, involving determining optimal solutions for mode and destination choice parameters, for input into strategic multi-modal transport models. These AI techniques have also been refined to ensure the model adheres to the Department for Transport’s Transport Analysis Guidance (TAG), a key user requirement.
Scope
Our project involved:
- Expanded Multi-Modal Model Testing: Development, testing and refinement of a model pipeline with existing model data.
- Demand AI Technique Refinement: Use of AI to develop a process for parameter prioritisation.
- End-to-End Integration: Integration with existing transport modelling tools as part of a wider model automation pipeline. This also involved accuracy analysis and quality assurance.
- AI Techniques for Transport Modelling Stages: Horizon scanning for further applications of AI-supported techniques for other transport modelling tasks.
- User Needs Research: Engagement with transport modellers, end model users and market competitors to further understand user types and their needs.
City Science Response
We proposed an AI-driven optimisation technique that aimed to reduce time and effort of manual processing in variable demand modelling.
Figure 1: Workflow for implementing AI techniques into our Transport Model
By employing AI optimisation techniques to identify optimal parameters, the tool consistently converges on the same solution each time. This contrasts with manual methods, where human operators may become stuck in less effective solutions (local minima).
In addition, transport models typically need to follow TAG whereby convergence criteria can be challenging to achieve. This means that if something changes in the model, like fuel costs, it should cause a predictable change in the resultant demand for car use based on known elasticties.
Outcomes
1. Reduced Effort: We have successfully leveraged AI capabilities to reduce the time and effort associated with manual processing of variable demand model calibration. This tool can be set running and then will require no human intervention until it has converged on the solution.
Figure 2: Demonstration of workflow of AI-optimisation techniques into our Variable Demand Model which then translates to assigned flows for public transport and highway networks
2. Improved Solution: Typically, a variable demand model could go through approximately 10 or 20 manual calibration iterations to achieve a convergence of between 0.2 and 0.4 for the demand-supply gap. Our findings show the process we have developed can reach a convergence of less than 0.1 in line with TAG requirements and can also be compliant with the TAG realism tests concerning elasticties for changes in fuel costs, public transport fares and journey times. However, further improvements and testing are required to understand the balance between computing power required and the quality of the solution, such that our tool is considered a viable alternative for future use.
We created Cadence visualisation from our model runs:
Figure 3: Visualisation of the demands on Cadence
Next Steps
We have identified that our tool is currently at Technology Readiness Level (TRL) 5, indicating that some integration in a relevant environment has been achieved. Our goal is to continue to improve and further integrate our variable demand model calibration tool, developing a prototype that we can consider making publicly accessible or marketing to transport model users.
