Not Another AI in Construction Post, But One That You Might Find Useful

AI? Machine Learning? Neural Networks?

We’re all familiar with the barrage of content surrounding ChatGPT and the "Top 10 AI apps to download," as well as numerous firms promising to solve our every challenge through proprietary AI solutions—at a steep monthly cost and with our data privacy at stake. AI innovations are everywhere, boasting remarkable capabilities that seem almost too good to be true.

Yet, as we navigate through these advancements, the construction industry faces two significant hurdles: a technology that is pioneering and searching for its place, alongside a market that is either unsure of how to leverage these tools or hesitant to adopt them due to fear or lack of understanding.

Last October, I embarked on a journey to explore the capabilities of neural networks firsthand, using publicly available Iowa housing data to predict prices and tumor data to determine malignancy. This experiment, conducted on my personal computer, served not only as a proof of concept, but also demonstrated the feasibility of running AI projects locally, without expensive offsite hardware. Through this process I've dove deeper into the mechanic of various uses of AI and Machine Learning Technologies. With this foundation, I hope to share how we can pragmatically utilize this portion of AI within the Construction Industry, using the resources and information that we already have.

Fine Tuning - Music and Neural Networks

Imagine you're in a recording studio, standing before a soundboard packed with knobs and sliders. You have the tune that you want to reach in your head, you just don’t know how to adjust the soundboard to get there. Each knob adjusts a specific element of the music—bass, treble, volume, etc., with the right combination producing the tune you have in your mind. When you finally figure out how each knob needs to be adjusted to match the tune you have in your head, you can use that same set up for similar tunes. In that way, Deep Neural Networks (DNNs) operate on a similar principle, albeit in the realm of data rather than sound.

At their core, DNNs are complex algorithms designed to mimic the neural pathways of our brains, capable of learning from vast amounts of data. The "knobs" in this context are the weights and biases within the network, fine-tuned during the training process. The training involves feeding the network with input data (the raw music) and adjusting these knobs to minimize the difference between the predicted output and the actual output (the tuning process) until the Neural Network can accurately predict or classify new, unseen data.

Finding the Function

The main role of a Deep Neural Network is to identify the underlying function between given inputs and outputs. In simpler terms, it learns the relationship or pattern connecting the data you feed it (input) to the results you're aiming for (output). This process involves iteratively adjusting the network's internal parameters based on the feedback from each prediction attempt, akin to fine-tuning a sound until it perfectly matches the intended tune.

Imagine if you tried to solve the classical number problem of determining what number comes next:

[3 , 4 , 5 ]

[5 , 9 , 10.3 ]

[12 , 12 , ___ ]

In order to solve these types of problems, you’ll look at the first two lines, and work out a pattern from there. One you determine the pattern, or the function in this case, you’ll then apply that function to solve the last problem.

In this case, the formula involved was simply the Pythagorean Theorem, whereby the solution to the last set was 17.

But computers don’t necessarily solve problems by referencing formulas like we do. They do what we were always told not to do in school: try every solution until it works. For simpler problems like the one above, a computer could try 1000’s of combinations in a matter of milliseconds. For more difficult problems, especially when adjusting the knobs in neural networks, they use optimization algorithms that effectively adjust the fine tuning to reduce training time.

(for more examples on optimization, reading up on genetic algorithms, in how they function better illustrates this kind of optimization, though not necessarily always tied to neural networks)

Practical Applications in Construction

What we ultimately have with Deep Neural Networks, is a way to take the data and knowledge that we already have, and look for patterns that we either cannot see, or don’t have the means to pour through what could be millions of lines on an excel sheet.

Lets look at a couple of examples of the data we have:

Lets say you’re bidding a AACE Class 3 Project with a Division 3 Cast in Place Concrete Foundation. It’s a budget bid, so you could gauge its cost based on historical data, but your data seems varied:

If your project was 7000sqft in Missouri in May, there would be no way to extrapolate what that might cost based on just the data here. You might have more information from each job that you could add in more columns, and more jobs to add to the rows, but with variances it wouldn’t change the outcome.

And that’s were Deep Neural Networks come into play.

Just like the soundboard, the neural network will adjust its internal knobs on the known data to make sure the output matches what is was supposed to be. If the first input was information from Job A, then it needs to output $8,800, if not, it needs to adjust its internal knobs, and try again. This is what is known as training the network.

Now bear in mind, the network in reality would be very inaccurate if only trained on 4 projects worth of data, just a person would be if tried on 4 items and asked about a 5 item they had never seen before.

If we trained on 100 projects, or 1000 projects, or more, the neural network will be better able to predict the cost, in the same way that we on site, can identify quantity of work, or unsafe scenarios, or quantify how much time it takes to perform a job. All by repetition. Because of this we start to see that while certain conditions on jobsites change with scope, as with many other factors, many of the tasks and activities that we perform largely stay the same.

Weather Example:

The same applies to the weather. The larger we build, the more expensive our contracts become, the more work is influenced by weather. When we have weather data available, we can plug that in a Deep Neural Network to predict how much of a risk we have:

Whether it’s per day or per hour, or over years, or has more columns, inputting these into a Deep Neural Network will give you a much better idea of what risk you have regarding weather. The quality of our data could, to an extent, tell us each day about how much of a weather-related delay that we could incur, and actually quantify that into a number before the contract is signed.

Other Examples:

-Chances of winning a bid based on previous submissions

-Leading Safety indicators, from PPE purchases and supplies on site

-Actual costs for construction site mobilization based on invoices.

-Links between punch list items on site, supply chain, subcontractor choice:

Be able to determine based on the hiring of certain contractors and subcontractors, what their potential punch item cost value would be based on previous work, and more effectively negotiating retainage.

-Weather and maintenance: based on the weather, how does that affect our maintenance cost?? Does it negatively affect our ability to perform works?

-Customer needs: What are the variety of requests and needs do our customers have? What are they asking for? What are they not asking for?

A properly tuned cost model for estimation is critically dependent on having the most up to date cost data to be effective in any bid. While that information is usually available in an organization, it is never where it needs to be, to be used, or is determined to be to much of an effort to incorporate it into a cost model. With Deep Neural Networks, coupled with proper database methods, and other data capture tools, you can dynamically update your cost models, with costs you are actually seeing, and be able to better see not just the high dollar items, but where every dollar goes.

Last Thoughts

As we’ve worked through examples of Deep Neural Networks and the types of problems they can solve, we can start to open the door to questions and problems we’ve determined to time consuming to solve, or thought we didn't have the relevant data. The inputs listed here are straight numbers from the data that we see day-to-day, but they can easily be converted to photos, audio or text, to solve a variety or other problems. These new tools will allow us to plan for things we couldn’t plan before. They will allow us to find patterns in how our sites and organizations function with each other. Ultimately they will allow for us to not only to better understand what we build, but how we can build better.

?

For a more detailed, very technical overview of how DNN’s work, I’d recommend this video from 3Blue1Brown’s YouTube Channel:

https://www.youtube.com/watch?v=aircAruvnKk&t=991s

Other Topics to read up on that will help the more technical understanding:

-Matrix Multiplication

-Back Propagation

-Gradient Descent

-And a good working knowledge of Python doesn't hurt.

There are many other Topics relating to the world of AI and Machine Learning including but not limited to: Image Classification, OCR, Vector Databases, LLMs, and much more that I hope to cover at a later date.