As known, autonomous vehicles have to run themselves in an optimal way (time-optimal, energy optimal…). With the rapid development of modern optimization techniques, the way to make vehicles smart and autonomous has been more and more.Take time-optimal problems as an example (going from point A to point B with least time spent), we have following available algorithms to specify the optimal control profiles (velocity profile or acceleration profile).
1. Genetic algorithms.
Different velocity combinations behave like genes, they belong to different chromosomes and they exchange genes with a certain rate of mutation. In this way numerous combinations can be found and tried on the model by computers until the one combination which makes time least is found.
2.Particle Swarm Optimization
Different combinations of velocities (the solution of the optimization problem) are represented as dots which are flying at certain velocity and altitude in 3D space. The velocity and altitude of each of these dots will be influenced by other dots, and the way they influence each other is towards the direction of optimization. Therefore, every velocity combination will be tried on the internal model until the optimal one found.
The common thing that these two algorithms share is that they have a mechanism of “mutation” to make sure the velocity combinations are diverse so that they can find a global optimal solution instead of a local one. With the same principle, there are more similar algorithms.
This one is really a part of artificial intelligence because it makes vehicles behave like human beings! Just like ourselves, we have the complete neuronetwork around our body so we can feel the surroundings and make decisions based on these feelings. And we tend to make more correct decisions with the feeling experience accumulated. Cars can be like this as well! We can teach them to feel the surroundings, to try different actions , and when they grow up, they will behave in a more and more smart way! Just like the growth of children!