I recently discovered the GP and find it fascinating. To explicitly learn the mathematical expression is quite different from the implicit learning neural networks do! I'm interested in how GP works with on-line learning, and I imagine since there is an evolutionary process on-line learning feels natural. But, I'm not sure I've quite grasped all the details, and it makes me question whether my intuition is right. So, I think if one feeds one data point at a time, the algorithm should update the mathematical expression, and perhaps a few data points at a time might yield better results. I think the max_samples parameter for gplearn allows me to specify what percentage of data points to look at once, but do all data points have to be available? What if all data points are not available? What would the loop below do?
While data keeps coming:
est_gp.fit(data[0], data[1])
Each time est_gp.fit is run, the method goes through N possible functions and modifies functions during each generation. However, if it does this for one data point, when a new one is introduced, will it take the winning model from the previous data point and throw it into the new population?
(Note: the same questions are posed on GitHub).
GP is quite strong and flexible.
As in other Machine Learning methods, all the data points should be available when you fit the model. The fitness function accounts for the current training set made available to the model. New data points can be added to your training data and then used to continue evolving.
You need to set the "warm_start" variable to true to continue learning from the current evolution state:
est_gp.set_params(generations=20, warm_start=True)
If you fit with only one data point, the evolution will have no reasons to keep the old models alive. The population will quickly converge to models that care about the new data point only.