A Reinforcement Learning Framework for Computing Eco-Driving Strategy
Eco-driving involves adaptively changing the speed of the vehicle to ensure minimal fuel consumption. We pose this problem within the framework of Markov decision problem with discounted reward. The key difficulty lies in identifying the state space and the reward function of the vehicle to be able to use reinforcement learning methods so that the vehicle can learn not only the optimal driving strategy, but also the rules of the road through reinforcement learning method. We use deep Q learning and DDPG to determine the optimal driving strategy.