Master Thesis about multi-agent Deep Reinforcement Learning for autonomous driving.
First Carla was transformed into an openAI Gym enverioment and then several architecture for both an encoder and a policy network were used.
First a VAE was used as an encoder but it wasn't powerfull enough to encode the input images and when the policy netwokr "looked" at the latent space it couldn't barely keep a vehicle in lanes.
Next an encoder using an RESNET was used an it performed way better, on the policy network side, it was a LSTM trained with SAC (altough other algos are implemented SAC performed better).
The main contribution of this work is a new algortihm composed SAC (Soft-Actor-Critic), A3C (making it distributed) and Geometric Deep Learning, by making the different agents (vehicles) a graph a message parssing algorithm is used.
General Diagram:
Encoder:
In order for the RL agents to perform better, the encoder needs human features which are added in the CarlaGymEnv.py before the step function that gives the input to the agent.
TODO: Change to separate file, like encoder.py
RL Diagram:
The algoritm uses as a base SAC (Soft-Actor-Critic).
