Giving Sight to Recombination for More Scalable Genetic Programming
Host
Una-May O'Reilly
ALFA Group, CSAIL
Abstract:
With modern society increasingly demanding AI systems to be explainable, Genetic Programming (GP) has huge potential for impact, because it can synthesize machine learning models as programs composed of human-interpretable instructions. While many machine learning algorithms operate by "training" (e.g., tuning weights by gradient descent), GP mostly operates by "searching", i.e., it iteratively and stochastically recombines the instructions of a population of programs into new programs, and discards the worst performing ones (survival of the fittest).
Unfortunately, this search approach can be computationally expensive, especially when traditional recombination methods are employed. Traditional recombination methods, which are also called "blind", make no attempt to harness problem information that potentially emerges during the search.
In my PhD, I worked on making recombination more "sighted", to make GP more scalable. I will first present the GP version of the Gene-pool Optimal Mixing Evolutionary Algorithm (GP-GOMEA), which attempts to learn on the fly what instructions belong together into building blocks and proceeds by recombining entire building blocks, to obtain a knock-on effect on scalability. I will then describe our work on assessing whether using GP for feature construction can contribute to make a secondary machine learning algorithm deliver more interpretable outcomes. I will finally describe how I improved Semantic Backpropagation-based GP, which performs recombination by leveraging program "semantic", i.e., the output, to be practically capable of scaling on realistic supervised regression problems.
Bio:
Marco Virgolin is a PhD candidate at Centrum Wiskunde & Informatica in Amsterdam, the Netherlands, and is enrolled at the Delft University of Technology, Delft, the Netherlands. He holds an M.Sc. in Computer Engineering from the University of Trieste, Italy. Marco is mostly interested in evolutionary and explainable machine learning, with a special focus on genetic programming. He has worked jointly with the Department of Radiation Oncology of the Amsterdam University Medical Centers-location AMC, to apply evolutionary machine learning in healthcare.
With modern society increasingly demanding AI systems to be explainable, Genetic Programming (GP) has huge potential for impact, because it can synthesize machine learning models as programs composed of human-interpretable instructions. While many machine learning algorithms operate by "training" (e.g., tuning weights by gradient descent), GP mostly operates by "searching", i.e., it iteratively and stochastically recombines the instructions of a population of programs into new programs, and discards the worst performing ones (survival of the fittest).
Unfortunately, this search approach can be computationally expensive, especially when traditional recombination methods are employed. Traditional recombination methods, which are also called "blind", make no attempt to harness problem information that potentially emerges during the search.
In my PhD, I worked on making recombination more "sighted", to make GP more scalable. I will first present the GP version of the Gene-pool Optimal Mixing Evolutionary Algorithm (GP-GOMEA), which attempts to learn on the fly what instructions belong together into building blocks and proceeds by recombining entire building blocks, to obtain a knock-on effect on scalability. I will then describe our work on assessing whether using GP for feature construction can contribute to make a secondary machine learning algorithm deliver more interpretable outcomes. I will finally describe how I improved Semantic Backpropagation-based GP, which performs recombination by leveraging program "semantic", i.e., the output, to be practically capable of scaling on realistic supervised regression problems.
Bio:
Marco Virgolin is a PhD candidate at Centrum Wiskunde & Informatica in Amsterdam, the Netherlands, and is enrolled at the Delft University of Technology, Delft, the Netherlands. He holds an M.Sc. in Computer Engineering from the University of Trieste, Italy. Marco is mostly interested in evolutionary and explainable machine learning, with a special focus on genetic programming. He has worked jointly with the Department of Radiation Oncology of the Amsterdam University Medical Centers-location AMC, to apply evolutionary machine learning in healthcare.