Skip to content

Heterogeneous Hierarchical Multi Agent Reinforcement Learning for Air Combat

Notifications You must be signed in to change notification settings

IDSIA/hhmarl_2D

Repository files navigation

HHMARL 2D

Heterogeneous Hierarchical Multi-Agent Reinforcement Learning for Air Combat Maneuvering, the implementation of the method proposed in this paper.

Overview

We use low-level policies for either fight or escape maneuvers. These will be first trained, then employed in the high-level hierarchy as part of environment.

Requiered Packages

  • ray["rllib"] == 2.4.0
  • torch >= 2.0.0
  • numpy == 1.24.3
  • gymnasium == 0.26.3
  • tensorboard == 2.13.0
  • pycairo == 1.23.0
  • cartopy >= 0.21.0
  • geographiclib == 2.0
  • tqdm

Training

Run train_hetero.py for heterogeneous agents training in low-level mode and train_hier.py to train the high-level policy (commander). The low-level policies must be pre-trained and stored in order to start training of the commander policy. At this stage, low-level policy training is configured for 2vs2 and high-level policy training for 3vs3. The reason for this is the structure of Ray for setting up Centralized Critics. But evaluations can be done in any combat configuration.

Procedure

For training the full model, proceed as follows:

1) Run train_hetero.py

  • Start with level=1 and agent_mode="fight". When finished, continue training up to level=4 with agent_mode="fight". It is important to stop at level 4, for now.
  • Run file with level=3 and agent_mode="escape".
  • Run file with level=5 and agent_mode="fight".
  • Finally, run file with level=5 and agent_mode="escape". This is not crucial, but recommended.

2) Run train_hier.py

  • Run file to train commander. No further steps needed.

3) Run evaluation.py

  • Set eval_hl=True to evaluate model with commander.
  • You can set any n-vs-m combat configuration now. For this, set the parameters num_agents and num_opps.
  • You can also change the opponent behavior with hier_opp_fight_ratio. This specifies the probability for choosing fight policy, i.e., hier_opp_fight_ratio=100 sets opponents purely in fight mode, and hier_opp_fight_ratio=0 purely to escape mode. Default is hier_opp_fight_ratio=75.
  • Set eval_hl=False to evaluate low-level policies without commander. Set also hier_opp_fight_ratio=100, otherwise you evaluate against fight and escape with the corresponding ratio.
  • With eval_hl=False, you can also specify which levels (3-5) you want to compare in any configuration. You can do this with eval_level_ag and eval_level_opp to set the levels together with num_agents and num_opps to specify the combat scenario.
  • Inside the evaluation folder, you get a .json file with all metrics.

Curriculum Learning

  • Training is done in levels (1-5) for fight policy. The opponent behaviour per level is as follows. L1: static, L2: random, L3: scripted, L4: uses L3 policy, L5: uses L3, L4 and escape policies.
  • The relevant arguments will be set automatically when training with the above procedure. But you can also manually set level, agent_mode and restore before starting training.
  • The algorithm checkpoint and the rendering images will be stored in results/L{X}_{mode}_2-vs-2.
  • High-level policy is not trained in curriculum fashion.

Configurations

Most important arguments to set are the following. All arguments can be found in config.py.

  • agent_mode is either "fight" or "escape"
  • level from 1 to 5 (only for low-level)
  • rew_scale to scale rewards. Default 1.
  • glob_frac is a float number for reward sharing between agents. Default 0.
  • restore either True or False, to restore training. When training in the above procedure, it will be automatically set to True when level>=2.
  • gpu either 0 or 1, to use gpu or not. Default 0.
  • num_workers is number of parallel samplers (threads). Default 4.
  • epochs number of training epochs. Default 10'000.
  • batch_size to adjust PPO training batch size. Default 2000.
  • eval either True or False, for having rendered images in log folder. Default True.
  • render either True or False, to visualize the current combat scenario. It stores iteratively the current combat situation as current.png file in log folder. When the file is opened in VS Code while the evaluation process runs, you get a "video" of the combat scene.
  • map_size is a float that will be mapped as -> x*100 = x[km], e.g. 0.3 -> 30 km per axis.

Inference

Levels 4 and 5 use the previously learned policies (fictitious self-play). Ray seems inconsistent when calling its method Policy.compute_single_action(). Therefore, the learned policies will be stored during training in folder policies from level 3 onwards. The actions will then be computed manually inside the method _policy_actions(). You can also manually export policies by running policy_export.py (have a look at it and make configurations as you want).

Commander Sensing

Change N_OPPS_HL in env_hier.py, train_hier.py and ac_models_hier.py to change detected opponents (N2-vs-N3 in the paper). E.g. setting N_OPPS_HL=3 allows the Commander to detect 3 opponents for an agent and can select one of these three to attack.

GPU vs CPU

Ray allows training on GPU but during several experiments, the performance was worse compared to CPU. Reason still unknown. This might improve in future versions. In our case, GPU was an RTX 3080Ti and CPU i9-13900H.

Note

HHMARL 3D is on its way with more advanced rendering ...

Citation

@misc{hhmarl2d,
  author = {Ardian Selmonaj and Oleg Szehr and Giacomo Del Rio and Alessandro Antonucci and Adrian Schneider and Michael Rüegsegger},
  title = {Hierarchical Multi-Agent Reinforcement Learning for Air Combat Maneuvering},
  year = {2023},
  eprint = {arXiv:2309.11247},
}