Constrained Generative Control and Mixed Autonomy

In a world of increasing automation, the need for algorithms to co-exist with humans is of great concern. On one hand, automation relieves human participation and in-creases efficiency and effectiveness. On the other hand, humans find joy in performing tasks currently being automated.

We present a generative deep learning methodology applied to the problem of controlwith human participation. We utilize the problem of autonomous driving as a problem to convey our approach. There are a variety of approaches to autonomous driving,each with their benefits and limitations. These range from difficulty training systems to models not having sufficient problem exploration. Our approach aims to consolidate these benefits while alleviating their limitations through the use of generative deep learning.

We demonstrate the ability to learn an entire distribution of driving behaviours from random walks, and select behaviour that results in good performance. Our method was tested on three road environments with contrasting human behaviours and showed the ability to perform well with human participation.