Our Final Aim was to learn a policy to perform lane change and merge for Connected Autonomous Driving using Graph Neural Networks in a Multi Agent setting .
We initially wanted to proceed with the environments given in Multi-Agent Connected Autonomous Driving using Deep Reinforcement Learning as mentioned in our project proposal. Since we found out highway environment is not present in the above mentioned paper. We went with another Open AI based simulator.We extended the single agent highway environment present in the gym to account for multi agent highway environment.
We adopted DQN as it was easier to interface with our simulation environment. We intend to use Stable baselines which is an Open AI gym based RL training library . But the problem is that it does not support Multi Agent training. So we are in the process of adopting it for Multi Agent training.
And we also plan to use our graph based Neural network architecture to show the improvement in policy performance with ablation studies.
Assumptions in our Present Work
Each agent knows the state of other agent during Training.
Policy is learnt in a centralized approach
Future Work
Policy to be trained in a decentralized approach employing state of the art algorithms such as PPO.
Employ training in different scenarios such as the presence of Aggressive and Conservative vehicle models.
Perform Ablation studies to analyze the Results obtained from different models.