xuxiyang1993 / MultiAgent-Guidance-Communication

MultiAgent-Guidance-Communication

A Python implementation of the algorithm proposed in paper "Multi-Agent Autonomous Operations in Urban Air Mobility with Communication Constraints"

Requirements

• python 3.6
• numpy
• gym
• ...

Getting Started

Make sure that you have the above requirements taken care of, then download all the files. You can run it using

cd MCTS
python Agent_vertiport.py

Optional arguments:

--save_path the path where you want to save the output

--seed seed of the experiment

--render set to True to visualize the simulator while running exp

--decentralized set it to True for decentralized control, default False

MCTS Algorithm

The whole MCTS algorithm files are under the directory MCTS\

common.py defines the general MCTS node class and state class

nodes_multi.py defines the MCTS node class and state class specifically for Multi Agent Aircraft Guidance proble, e.g., given current aircraft state and current action, how to decided the next aircraft state

search_multi.py describes the search process of MCTS algorithm

Simulator

The simulator codes are under Simulators\

config_vertiport.py defines the configurable parameters of the simulator. For example, airspace width/length, number of aircraft, scale (1 pixel = how many meters), conflict/NMAC distance, cruise/max speed of aircraft, heading angle change rate of aircraft, number simulations and search depth of MCTS algorithm, vertiport location, ...

MultiAircraftVertiportEnv.py is the main simulator. Some main functions include:

• __init__() initilize the simulator by generating vertiports, centralized controller, loading configuration parameters, generating aircraft.

• reset() will reset the number of conflicts/NMACs to 0 and reset the aircraft dictionary. Note here all the aircraft objects are stored in the AircraftDict class, where we can add/remove aircraft from it, and get aircraft object by id.

• _get_ob() will return the current state, which is n by 8 matrix, where n is the number of aircraft. Each aircraft has (x, y, vx, vy, speed, heading, gx, gy) state information.

• _get_normalized_ob() will return the normalized current state, which will potential be useful if we want to feed state into a neural network.

• step() will return next state, reward, terminal, info given current state and current action. Each aircraft will fly according the given action. We have a clock at each vertiport to decide whether to generate new flight request/aircraft.

• _terminal_reward() will return the reward function for current state. This function will check if there is any conflict/NMAC between any two aircraft and update conflict/NMAC number. It will also remove aircraft that reaches goal position and aircraft pair that has NMAC.

• render() will visualize all of the current aircraft and vertiport. I used this to generate the demo video.

• Controller class is the centralized controller, which can receive information from aircraft and process the missing information.

If you have any questions or comments, don't hesitate to send me an email! I am looking for ways to make this code even more computationally efficient.

Email: [e-mail]