General Principles of Machine Learning

Think of a machine learning model as a "consultant" you hand your experiment results to. It analyses them, finds the underlying trends, correlations, sensitivities, etc. Once it has acquired enough data, you can then ask it what it thinks of potential new experiments you don't yet have results for. It uses the data and the understanding it has acquired to tell you what result it expects.

The most basic machine learning model family is linear regression: you have a set of experiment results (a scatter of data points), you construct a line that passes through them (one that has the least cumulative distance to all of the points) and you consider that the results of unknown experiments fall on that line.

The model here is the line you have constructed. One might ask: why not a hyperbola? A logarithm, or an exponential function? These are very relevant questions.

Some people say that model building is more of an art than it is a science. There is no general rule for deciding which model to go with in which case. It all depends on the judgment of the statistician. In some cases, linear models are more than enough. In others, they are not.

Uncertainty of predictions

Let's say you have built a linear regression model. When you predict some outcome using your model, you get a result. But you don't know how certain that specific prediction is.

The most you can get is a mean-squared error, which tells you about the overall quality of the model, not the error of a single prediction.

Some of the quite advanced machine learning model families work like this: they give you overall quality measures, but not the uncertainty of a single prediction. E.g. Random Forests, Neural Networks, etc.

However, there are some non-parametric model families that do. That is, each prediction comes with a confidence interval. The shorter the interval, the more certain the prediction is.

How ML is implemented in Alfraido

Alfraido relies on a state-of-the-art non-parametric model that evaluates uncertainty at each prediction. You tell it the parameters of your material and it predicts the performance outcomes you are likely to have when you test, each one along with its own confidence interval.

This matters for two reasons. First, you know which predictions to trust: a narrow interval means you can rely on the number, while a wide one is a signal that the model is still guessing and that this region deserves a real experiment.

Second, and more importantly, this uncertainty is what lets Alfraido tell you which experiment to run next. Instead of testing blindly, the model weighs where the best outcomes are likely to be against where it is still unsure, and points you toward the experiments that teach it the most while moving you closer to your target. That is how you reach your goal in far fewer experiments than a classical trial-and-error approach.