Evolving recurrent neural networks for emergent communication
Joshua Sirota, Vadim Bulitko, Matthew Robert Graham Brown, Sergio Poo Hernandez · Proceedings of the Genetic and Evolutionary Computation Conference Companion · 2019
Recent research showed that deep neural networks can be trained to create shared languages to communicate and cooperate with each other. These approaches used fixed, handcrafted network architectures which were trained with reinforcement learning. We extend this approach by using neuroevolution to automate network design and find network weights of communicating agents. We show that neuroevolution is a viable approach for training agents to develop novel languages so as to communicate amongst themselves.