Scientists from the Vrije Universiteit Brussel’s AI Experience Center have developed a new algorithm for an autonomous wheelchair. The algorithm learns faster and independently on the basis of trial-and-error. This should prevent risks such as collisions with obstacles and people, making self-driving wheelchairs more comfortable and safe.
The control algorithm uses reinforcement learning, a form of machine learning whereby the computer learns by experimenting in the environment in which it should be active. Classical reinforcement learning requires many interactions, which lengthens the learning process. The new algorithm can reduce the learning time to 1-2 hours.
Principal investigator Denis Steckelmacher explains: “Reinforcement learning is actually a system based on trial-and-error. In order to speed up the learning process, our algorithm cleverly combines different types of reinforcement learning. In this way, one compensates for the weaknesses of the other. This results in a learning process that’s very quick if we’re sufficiently sure, and carefully adjusts when there’s uncertainty.”