Causal ML: AI/ML That Actually Works in Deployment

July 1, 2024

Prof. Daniel Franks

Why is it that despite all advances in AI and ML, we still see so many real world AI deployments fail? According to Gartner, about 85% of Artificial Intelligence deployments fail because standard AI is not designed for application in the real world.

The problem is that standard approaches are designed purely for predicting based on historical data. They’re great for prediction if the future looks like the past, but if they’re used to decide on real world actions to shape outcomes, the future does not look like the past. In fact, the whole purpose is to shape the future in our favour. This is a prediction-upon-action problem, not just a prediction problem. Only Causal AI can do this.

Not only should the deployment of AI solutions be in terms of predicting outcomes, but also with respect to the effect of interventions. It is for precisely this reason that causal AI has been designed. Unlike the standard AI, which is remarkably great at pattern recognition, causal AI models cause-and-effect relations between variables, thereby greatly making it more robust in terms of real-world applications.

In this blog, we will go through a simple-demo using our flagship platform, CausaDB, to illustrate the contrast between traditional AI versus Causal AI. We will take a use case for the optimization of heating control systems in commercial buildings to show why Causal AI is necessary for effective deployment.

Demo Overview: Optimizing Heating Control Systems

Our CausaDB platform embodies an industry agnostic platform for Causal AI, but we will work with an example of ML for optimizing heating control though HVAC. For this demo we are going to work with a simulated dataset representative of real-world deployment:

This way, we can see what happens in deployment. In this simplified dataset, we have variables such as the day, outdoor temperature, indoor temperature, and overall energy usage along with settings of the HVAC heating—the key variable we are trying to optimize.

Traditional AI Approach

Let’s start by using a traditional AI/ML approach with the popular XGBoost algorithm. We’re using this because of its popularity, but our results apply equally to any non-causal approach. Using XGBoost we predict heating settings based on outdoor and indoor temperatures. You can see the code in our code example: https://www.docs.causadb.com/docs/examples/use-cases/causadb_heating.

The model performs strongly using the industry standard validation approach of assessing it on unseen historical data:

This is normally the point where the algorithm would be seen fit to be deployed. However, let’s look at how it does when deployed. To do this we generate another year of data for the outdoor temperature and try to control the indoor temperature by inputting a target of 18 degrees and taking its HVAC setting recommendation for each day. The result of this is:

Test data
Performance of standard AI (target temp 18 deg)

You can see that it performs badly. It’s around two degrees off and highly unstable. This raises a critical flaw: even with high test accuracy, standard AI tends to fail in real-world deployment.

Causal ML with CausaDB

Now let’s turn to CausaDB. As before, you can find the code in our code example. Now let’s compare the deployment performance of CausaDB with standard AI. The results are striking: CausaDB strongly outperforms XGBoost. The savings from this would are massive.

CausaDB actually works in deployment

Conclusion

Causal AI is a necessity for any real-world deployment where businesses are looking for AI recommendations on actions, or to automate actions. Standard AI will fail on this, while CausaDB provides dependable and optimal actions to drive substantial benefit. This principle applied across all industries, where optimal ML informed actions are desired.