Machine Learning

The Elara platform can merge together machine learning with simulation and optimization. In order to use machine learning in a simulation, ProcessBuilder allows ML functions to be injected into processes. Each injected ML function makes a prediction based on input features (and the trained ML model) to produce an East value (like a float or a string).

Note that some ML functions incorporate randomness. That is, the ML function predicts a statistical distribution given some features (as a Bayesian process). Evaluating such an ML function within ProcessBuilder will draw a random sample from this distribution during the simulation. The predicted value may differ from trajectory to trajectory. The scheme used for seeding the random number generator on the Elara platform was detailed in the previous tutorial.

ML expressions

An ML expression is a special expression type that can only be used in ProcessBuilder. They are invoked by providing a set of features required by the ML model to perform evaluation. It will return a prediction based on the pre-trained machine learning model. (The quality of the prediction will depend on the quality of the model, training data, training process, etc.)

new ProcessBuilder("example")
    .ml(predict_a_float)
    .value("feature1", IntegerType)
    .value("feature2", StringType)
    .let(
        "prediction",
        (props, resources, mls) => mls.predict_a_float(Struct({
            feature1: props.feature1,
            feature2: props.feature2,
        }))
    )

Note the usage of Struct to provide the names and values of the various ML features.

Such ML expressions are always injected by ProcessBuilder and cannot be constructed directly by the user.

Next steps

Continue to the next tutorial to understand how to mutation works in imperative contexts in East and Elara.