Design and Training of AI Agent using Deep Q - Learning and Carla

Abstract

The construction of artificial intelligence agents for controlling the vehicles autonomously is computationally expensive and requires more hardware resources in training the agent. This research focuses on constructing a low-power agent which can be deployable in space-constrained devices. Carla is an open-source environment, that gives a simulation of real-world experiences especially in training the artificial agent with stochastic elements which is similar to the real human world. As the real-world training of the agent can be costly and requires more resources, simulation helps in creating an environment with fewer resources and results in better accuracy in the real-world deployment of the model. Deep Q - Learning is a form of reinforcement learning algorithm, that is used as the base for creating the agent. The DQL is an advanced version of Q - Learning which helps in handling the memory in training and testing the agent, as the Q - Learning requires a lot of memory.