< Back to previous page


SHE for HIM: Simulated Human Environment for the Human Interface Mate

Machine learning is giving computers the ability to learn, or adapt their behavior, based on information they receive. By giving these "learning programs" the right information, they can be trained for a specific purpose. At this point in time, most of the machine learning algorithms that are trained need a large amount of training data in order to become good at something. One could train a program to recognize cars in a picture for example, by showing it a thousand pictures, and telling it for every picture where there are cars to be seen. Preparing this training data would consist of gathering a thousand pictures of cars, and labelling every car in each of the pictures. A possible alternative to doing this manually, is to generate life-like pictures of cars using another computer program. The labels on the pictures, marking every car, can be generated as well. These generated pictures could then be fed as training data to the machine learning program. Using this synthetic data generation to train machine learning programs has already been proven to work on real applications, with good results. This research aims to first apply this technique to a real industrial application or process. After that, devising a feedback method used for combining both synthetic and real process data to accelerate and improve learning will be attempted. The research will also explore neuro-evolution techniques for accelerated and better training. 

Date:1 Oct 2018  →  Today
Keywords:neuro-evolution, machine learning, domain randomization, synthetic data generation
Disciplines:Ceramic and glass materials, Materials science and engineering, Semiconductor materials, Other materials engineering
Project type:PhD project