End-to-end Control of Kart Agent with Deep Reinforcement Learning
Ruiming Zhang, Liu Chengju, Chen Qijun · 2018
Research on kart racing games provides an opportunity to develop real autonomous driving controllers using unsupervised learning methods and therefore worth studying. In this paper we bring forward a feasible end-to-end control technique using a novel DQN architecture. Recurrence is introduced to eliminate the effect caused by the partially observable environment. A weighting layer is created to strengthen the impact differences of the lengthy episode. A modified relay mechanism is applied to solve the problem of sparsity and speed up learning. The method is tested and verified under the Mario Kart64 environment. With the screen pixels being the only input during both training and testing period, a direct speed control signal being the output of the system, this method makes the agent perform out similar learning behaviours just like a human player. The resulting agent is able to recognize the environment features well and could act out advanced behaviours which are undiscovered in previous methods.