From AI-driven forecasting to rapid decision making: why we built Driving Paths

It’s been nearly 10 years since I wrote the first line of code that set me on the path of Machine Learning (ML) algorithms, and I still believe that results derived from ML models are not easy to understand. 

There. I said it.

These models deal in the realms of predictions and probabilities - concepts that most people don’t interact with on a daily basis. A company that uses Machine Learning must also make sure that their users have a healthy interaction with the results from their models.

In the past five years at nPlan I have been lucky enough to work on the deployment of ML-powered products in an industry that has seen little adoption of AI technologies in general: management of large capital projects. In this post I would like to share one success story and why I think it worked.

One of the biggest time-sinks in project management decision-making is...

...the process of quantitative risk analysis. It involves estimating probability distributions for delays to individual activities and (drumroll) running a Monte Carlo simulation. The “traditional”* way to do this can take days, if not weeks, of manual work. At nPlan we have made this process take minutes by using an ML model(s) that can estimate the probability of delay from historical data.

With this model, we are able to estimate the end dates of any construction project, regardless of size**, in a fraction of the time. So, problem solved, money rolls in, job done, … right? Well, the issue is that - forgive me if you saw this coming - an ML model is not enough to get people to make better decisions. Increasing velocity and predictive performance is only the first step in building a world where construction professionals can leverage their forecasts to improve their outcomes.

The development of Driving Paths

To create a full solution that delivers value we focused on the problem of determining (with confidence) the critical path for the project. The critical path is the ‘route’ through certain activities which, if blocked, will cause the project completion date to slip backwards***. 

In a deterministic world, the critical path is clear to everyone, but when you add uncertainty to the end dates of every activity in the project the picture becomes blurry. In the latter case, there are different paths through the project that could become critical depending on where delay happens.

Understanding which are the most likely critical paths is critical for management decision-making. Managers will want to spend most time and resources on the areas of the schedule that, if delayed, will most affect on-time completion. So how do you identify all these competing paths? In particular for larger projects (think 10k-50k activities), putting all these paths in a single visual is unimaginable.

This is where nPlan’s data scientists had a brilliant idea. They developed a (proprietary) algorithm called ‘Driving Paths’. Driving Paths summarises the schedule using the forecasted delays as a grounding mechanism. The algorithm results in a schedule graph that contains only the paths in the schedule that are most likely to both experience delay (as determined by the ML model) and to affect the end date (as determined by the schedule logic). All other paths through the schedule are discarded, thus resulting in a visualisation that is both interpretable and useful.

A rendering of driving paths for a schedule of 2000 activities is shown below. Each circle is an activity, each line is a connection between two activities and the thickness of the lines represents the likelihood that this path will be the critical path. Add some UI interactions and magic happens. You can quickly extract information that would otherwise take hours or days and a few people to extract.

Using driving paths, our users are capable of observing which sections of the schedule require the most attention. They have a few levers that they can pull to make better management decisions:

• They can see the likelihood that their planned critical path will indeed be the critical path. This can justify diverting management resources to (or from) it if other parts of the schedule are more likely to cause slippage
• They can manage several competing critical paths that have similar probabilities; this way, they can reduce the overall risk of slippage to the project end-date
• They can use this information to re-plan sections of the schedule so that complex portions (which would be hard to deliver) occur away from the likely critical paths

All of this is made possible by looking at a single visualisation that has been designed with the specific purpose of enabling better decision making. How is this possible?

What really matters…

…is not that there is an AI model sitting under the hood. We have all seen our fair share of AI hype that leads to an avalanche of AI applications with a new technology. There is an initial uptake, but usually when the dust settles, few people end up using the application meaningfully in their day-to-day lives.

What I have found really matters - and this will not come as a surprise to many - is whether a user can do something valuable using the outputs of the model. In the space that nPlan works, we focus on helping our users make better decisions about their projects, in a fraction of the time, than what is possible with the technologies that they currently have available. Driving Paths helps our users achieve exactly that.

To conclude

We know that this works because Driving Paths is one of our most popular features. Our users often cite it as one of the tools that are key to delivering nPlan results to various stakeholders because it delivers a key piece of information in a visualisation that is easy to understand.

Users are now leveraging their forecasts to improve the outcomes of their projects and they are making better decisions using Driving Paths. Nobody is talking about the AI model anymore… but none of it would have been possible without it.

Interested in learning more about Driving Paths or would you like to see a demo? Get in touch!

Footnotes

* I deliberately put “traditional” in quotation marks here because the world of risk estimation in construction is changing… fast.

** Yes, regardless of size, we have worked with projects that have 10s of thousands of activities and our forecasts are ready in minutes.

*** For the project control purists out there, I know that there is a more formal definition of the critical path. But trying to keep it light here.