Efficient Social Attention for Autonomous Decision-Making in Varing Traffic Scenarios

This repository contains the source code for coursework paper: Efficient Social Attention for Autonomous Decision-Making in Varing Traffic Scenarios. The experiment environment is forked from: HighwayEnv and rl-agents. We modify the code to deploy our proposed method.

Run the Code

To run the code, you may follow the instruction in .\example.ipynb.

Modified Code Files

The code files we modified are as follows, which contain the implementation of our proposed method.

├── rl-agents
│   ├── rl_agents
│   │   ├── agents
│   │   │   ├── common
│   │   │   │   ├── models.py -> introduce sparse attention into the model
│   │   │   ├── deep_q_network
│   │   │   │   ├── pytorch.py -> introduce learnable weight for double q learning
│   ├── scripts
│   │   ├── configs
│   │   │   ├── HighwayEnv
│   │   │   │   ├── env_obs_attention.json          -> Configure the environment observation state
│   │   │   │   ├── agent
│   │   │   │   │   ├── DQNAgent
│   │   │   │   │   │   ├── ego_attention_2h.json   -> Set up the agent's algorithm model
│   │   │   ├── IntersectionEnv
│   │   │   │   ├── env_obs_attention.json
│   │   │   │   ├── agent
│   │   │   │   │   ├── DQNAgent
│   │   │   │   │   │   ├── ego_attention_2h.json
│   │   │   ├── MergeEnv
│   │   │   │   ├── env_obs_attention.json
│   │   │   │   ├── agent
│   │   │   │   │   ├── DQNAgent
│   │   │   │   │   │   ├── ego_attention_2h.json
│   │   │   ├── RoundaboutEnv
│   │   │   │   ├── env_obs_attention.json
│   │   │   │   ├── agent
│   │   │   │   │   ├── DQNAgent
│   │   │   │   │   │   ├── ego_attention_2h.json

Results

The raw data and visualisations of the experiments can be found in the .\experiments folder and the demonstration videos can be found in the .\videos folder.

License

Code: MIT License