AI in an EV World

Artificial Intelligence (AI) in our Earned Value (EV) World 

 

Warning: Stop reading unless you are ready to do delve into the AI world of custom Neural Networks (NNs) and how AI is converging with our world of Project Management and Earned Value (EV). 

The first step in implementing AI is knowing which problems it can solve. Think of AI like auto-complete in your Google searches: it’s best for making predictions, finding irregularities in data, and offering suggestions. AI should not be used for problems with clear checks and corrections; those are better suited for algorithms. AI shines when you need to predict something based on new information, like assessing resumes or guessing future outcomes. 

Some AI Use Cases 

One good AI implementation detects if there is an irregularity within your current data. You can also use AI to read your existing EV or scheduling data and then try to predict the probability of staying on course, where it might go off course—and by how much. Another good AI implementation matches field types. Often, when you are trying to match a field type to another during your import-export-import dance, you use some tool that checks existing cases and maps them to known solutions. The problem with these types of fixes is that a new case always crops up and you then need to add a new case solution. You can feed AI these cases and conditions, and then when it encounters a new field type it has never seen, it can guess what the solution should be. This guessing capability is a much more stable solution to this type of problem.  

Key takeaway: Remember that none of these situations need to be general solutions. Every one of these AI implementations can be fed and trained on your company or project-specific data and give your own brand of custom solutions. 

Getting started: Goodbye 80/20 rule, hello 70/30! 

To get started, build your own custom neural network. Sounds intimidating, but there are many good sources to help put something together. The simplest path is to use Python’s Tensor Flow library to build a model and then train it. The art to building a well-trained model has a lot to do with trying to avoid overfitting. This is when you train your model so well on your existing data that it can reliably predict with incredible accuracy. When this happens, your model will be less able to correctly evaluate data it has never seen before. You typically want to use 70% of your data for the training and 30% for the validation.  

Models, Neurons, Layers, and the Power of 2 

There are a few things to consider within your model’s structure. Each element you can adjust when building your neural network ends up being more of an art than an exact science. You will need to do some experimenting to find the right balance. The first is the number of layers and number of neurons in each layer. You can start with any number of neurons, but try not to exceed 2 neuron layers. Typically pick some power of 2: 16, 32, 64, or 128. If your model seems to be overfitting , reduce the number of layers to 1. If It’s still happening, reduce the number of neurons.  

Epochs, Batch Sizes, and the Power of 2 (again) 

You will also want to determine how many epochs you want and your batch size. Your epochs are how many times your model will view some data and attempt to adjust itself. The batch size is how many elements from your data it will grab during each epoch. The number of epochs you should run is largely based on the size of your data. A good way to determine your epoch size is to set some number equal to the square root of your data size and see if you start getting diminishing returns of corrections. If you are, then decrease the amount of epochs to prevent overfitting. If you are not, increase the number of epochs. You will also want to experiment with batch size. It’s a good rule of thumb to choose a batch size that is also a power of 2, such as 32, 64, 128, or 256.  

AI is always, sometimes, maybe…wrong—or right? 

Anywhere you would use an AI solution, you must implicitly accept that the answer could potentially be incorrect. At its heart, you must train your AI on data or context for it to build a profile from which it can make smarter guesses. Never forget that these are still just guesses. You can build very good models that will be correct almost every time; however, there’s always the chance that their prediction, or result, is incorrect.  

Don’t let this discourage you!  

There are two important ideas to keep in mind when trying to weigh using AI. Firstly, if you compare a competently built AI model’s predictions to a human’s, you’ll find that the AI makes far fewer errors. Secondly, the evaluation time between the AI and a human is incomparable. What this means is that you will have different but fewer errors with AI, the operational time cost is way lower, and the human component can spend some of that saved operational time validating the result. In every case, reviewing a product is far less time consuming than creating a product. This allows AI to act as a large efficiency booster. 

Lets Chat 

This is a new frontier. We understand It’s complicated. This article was written to give you a good foundation into this new AI landscape. If you have other cool ideas to share or need some help on your AI and EV journey, please reach out. We are passionate about the future of custom neural networks for new EV solutions.  

---