Creating a Traffic Merging Behavior Using NeuroEvolution of Augmenting Topologies

One of the main goals in developing an autonomous vehicle is programming the action of merging into the traffic lane from an entrance ramp. We seek to create such a behavior through the use of NeuroEvolution of Augmenting Topologies (NEAT) by evolving an agent over many generations to maximize a certain prescribed fitness function, which encourages a smooth merging behavior without crashing. Our experiment environment is simulated, and the agent starts from a fixed position on the entrance ramp and tries to merge into a separate lane with a constant flow of traffic. After evolving on many generations where the frequency of traffic on the highway is constant, say k, we have created an agent that can merge into the traffic lane seamlessly at the frequency of k, but only at k. Moreover, by training the agent on multiple environments where the frequency of the traffic varies within a single generation of NEAT, we have created an agent that can merge onto the highway given any frequency of traffic, as long as it is constant throughout the run and light enough so that the agent can physically fit in. There are, however, still limitations to this merging behavior, and we will point to some potential future research areas.