ml-agents/docs/Training-ML-Agents.md

Training ML-Agents

The ML-Agents toolkit conducts training using an external Python training process. During training, this external process communicates with the Academy object in the Unity scene to generate a block of agent experiences. These experiences become the training set for a neural network used to optimize the agent's policy (which is essentially a mathematical function mapping observations to actions). In reinforcement learning, the neural network optimizes the policy by maximizing the expected rewards. In imitation learning, the neural network optimizes the policy to achieve the smallest difference between the actions chosen by the agent trainee and the actions chosen by the expert in the same situation.

The output of the training process is a model file containing the optimized policy. This model file is a TensorFlow data graph containing the mathematical operations and the optimized weights selected during the training process. You can set the generated model file in the Behaviors Parameters under your Agent in your Unity project to decide the best course of action for an agent.

Use the command mlagents-learn to train your agents. This command is installed with the mlagents package and its implementation can be found at ml-agents/mlagents/trainers/learn.py. The configuration file, like config/trainer_config.yaml specifies the hyperparameters used during training. You can edit this file with a text editor to add a specific configuration for each Behavior.

For a broader overview of reinforcement learning, imitation learning and the ML-Agents training process, see ML-Agents Toolkit Overview.

Training with mlagents-learn

Use the mlagents-learn command to train agents. mlagents-learn supports training with reinforcement learning, curriculum learning, and behavioral cloning imitation learning.

Run mlagents-learn from the command line to launch the training process. Use the command line patterns and the config/trainer_config.yaml file to control training options.

The basic command for training is:

mlagents-learn <trainer-config-file> --env=<env_name> --run-id=<run-identifier> --train

where

• <trainer-config-file> is the file path of the trainer configuration yaml.
• <env_name>(Optional) is the name (including path) of your Unity executable containing the agents to be trained. If <env_name> is not passed, the training will happen in the Editor. Press the ▶️ button in Unity when the message "Start training by pressing the Play button in the Unity Editor" is displayed on the screen.
• <run-identifier> is an optional identifier you can use to identify the results of individual training runs.

For example, suppose you have a project in Unity named "CatsOnBicycles" which contains agents ready to train. To perform the training:

1. Build the project, making sure that you only include the training scene.
2. Open a terminal or console window.
3. Navigate to the directory where you installed the ML-Agents Toolkit.
4. Run the following to launch the training process using the path to the Unity environment you built in step 1:

mlagents-learn config/trainer_config.yaml --env=../../projects/Cats/CatsOnBicycles.app --run-id=cob_1 --train

During a training session, the training program prints out and saves updates at regular intervals (specified by the summary_freq option). The saved statistics are grouped by the run-id value so you should assign a unique id to each training run if you plan to view the statistics. You can view these statistics using TensorBoard during or after training by running the following command:

tensorboard --logdir=summaries --port 6006

And then opening the URL: localhost:6006.

Note: The default port TensorBoard uses is 6006. If there is an existing session running on port 6006 a new session can be launched on an open port using the --port option.

When training is finished, you can find the saved model in the models folder under the assigned run-id — in the cats example, the path to the model would be models/cob_1/CatsOnBicycles_cob_1.nn.

While this example used the default training hyperparameters, you can edit the training_config.yaml file with a text editor to set different values.

Command Line Training Options

In addition to passing the path of the Unity executable containing your training environment, you can set the following command line options when invoking mlagents-learn:

• --env=<env>: Specify an executable environment to train.
• --curriculum=<file>: Specify a curriculum JSON file for defining the lessons for curriculum training. See Curriculum Training for more information.
• --sampler=<file>: Specify a sampler YAML file for defining the sampler for generalization training. See Generalization Training for more information.
• --keep-checkpoints=<n>: Specify the maximum number of model checkpoints to keep. Checkpoints are saved after the number of steps specified by the save-freq option. Once the maximum number of checkpoints has been reached, the oldest checkpoint is deleted when saving a new checkpoint. Defaults to 5.
• --lesson=<n>: Specify which lesson to start with when performing curriculum training. Defaults to 0.
• --num-runs=<n>: Sets the number of concurrent training sessions to perform. Default is set to 1. Set to higher values when benchmarking performance and multiple training sessions is desired. Training sessions are independent, and do not improve learning performance.
• --num-envs=<n>: Specifies the number of concurrent Unity environment instances to collect experiences from when training. Defaults to 1.
• --run-id=<path>: Specifies an identifier for each training run. This identifier is used to name the subdirectories in which the trained model and summary statistics are saved as well as the saved model itself. The default id is "ppo". If you use TensorBoard to view the training statistics, always set a unique run-id for each training run. (The statistics for all runs with the same id are combined as if they were produced by a the same session.)
• --save-freq=<n>: Specifies how often (in steps) to save the model during training. Defaults to 50000.
• --seed=<n>: Specifies a number to use as a seed for the random number generator used by the training code.
• --env-args=<string>: Specify arguments for the executable environment. Be aware that the standalone build will also process these as Unity Command Line Arguments. You should choose different argument names if you want to create environment-specific arguments. All arguments after this flag will be passed to the executable. For example, setting mlagents-learn config/trainer_config.yaml --env-args --num-orcs 42 would result in --num-orcs 42 passed to the executable.
• --base-port: Specifies the starting port. Each concurrent Unity environment instance will get assigned a port sequentially, starting from the base-port. Each instance will use the port (base_port + worker_id), where the worker_id is sequential IDs given to each instance from 0 to num_envs - 1. Default is 5005. Note: When training using the Editor rather than an executable, the base port will be ignored.
• --slow: Specify this option to run the Unity environment at normal, game speed. The --slow mode uses the Time Scale and Target Frame Rate specified in the Academy's Inference Configuration. By default, training runs using the speeds specified in your Academy's Training Configuration. See Academy Properties.
• --train: Specifies whether to train model or only run in inference mode. When training, always use the --train option.
• --load: If set, the training code loads an already trained model to initialize the neural network before training. The learning code looks for the model in models/<run-id>/ (which is also where it saves models at the end of training). When not set (the default), the neural network weights are randomly initialized and an existing model is not loaded.
• --no-graphics: Specify this option to run the Unity executable in -batchmode and doesn't initialize the graphics driver. Use this only if your training doesn't involve visual observations (reading from Pixels). See here for more details.
• --debug: Specify this option to enable debug-level logging for some parts of the code.
• --multi-gpu: Setting this flag enables the use of multiple GPU's (if available) during training.
• --cpu: Forces training using CPU only.

Training Config File

The training config files config/trainer_config.yaml, config/sac_trainer_config.yaml, config/gail_config.yaml and config/offline_bc_config.yaml specifies the training method, the hyperparameters, and a few additional values to use when training with Proximal Policy Optimization(PPO), Soft Actor-Critic(SAC), GAIL (Generative Adversarial Imitation Learning) with PPO, and online and offline Behavioral Cloning(BC)/Imitation. These files are divided into sections. The default section defines the default values for all the available training with PPO, SAC, GAIL (with PPO), and offline BC. These files are divided into sections. The default section defines the default values for all the available settings. You can also add new sections to override these defaults to train specific Behaviors. Name each of these override sections after the appropriate Behavior Name. Sections for the example environments are included in the provided config file.

Setting	Description	Applies To Trainer*
batch_size	The number of experiences in each iteration of gradient descent.	PPO, SAC, BC
batches_per_epoch	In imitation learning, the number of batches of training examples to collect before training the model.	BC
beta	The strength of entropy regularization.	PPO
demo_path	For offline imitation learning, the file path of the recorded demonstration file	(offline)BC
buffer_size	The number of experiences to collect before updating the policy model. In SAC, the max size of the experience buffer.	PPO, SAC
buffer_init_steps	The number of experiences to collect into the buffer before updating the policy model.	SAC
epsilon	Influences how rapidly the policy can evolve during training.	PPO
hidden_units	The number of units in the hidden layers of the neural network.	PPO, SAC, BC
init_entcoef	How much the agent should explore in the beginning of training.	SAC
lambd	The regularization parameter.	PPO
learning_rate	The initial learning rate for gradient descent.	PPO, SAC, BC
max_steps	The maximum number of simulation steps to run during a training session.	PPO, SAC, BC
memory_size	The size of the memory an agent must keep. Used for training with a recurrent neural network. See Using Recurrent Neural Networks.	PPO, SAC, BC
normalize	Whether to automatically normalize observations.	PPO, SAC
num_epoch	The number of passes to make through the experience buffer when performing gradient descent optimization.	PPO
num_layers	The number of hidden layers in the neural network.	PPO, SAC, BC
pretraining	Use demonstrations to bootstrap the policy neural network. See Pretraining Using Demonstrations.	PPO, SAC
reward_signals	The reward signals used to train the policy. Enable Curiosity and GAIL here. See Reward Signals for configuration options.	PPO, SAC, BC
save_replay_buffer	Saves the replay buffer when exiting training, and loads it on resume.	SAC
sequence_length	Defines how long the sequences of experiences must be while training. Only used for training with a recurrent neural network. See Using Recurrent Neural Networks.	PPO, SAC, BC
summary_freq	How often, in steps, to save training statistics. This determines the number of data points shown by TensorBoard.	PPO, SAC, BC
tau	How aggressively to update the target network used for bootstrapping value estimation in SAC.	SAC
time_horizon	How many steps of experience to collect per-agent before adding it to the experience buffer.	PPO, SAC, (online)BC
trainer	The type of training to perform: "ppo", "sac", "offline_bc" or "online_bc".	PPO, SAC, BC
train_interval	How often to update the agent.	SAC
num_update	Number of mini-batches to update the agent with during each update.	SAC
use_recurrent	Train using a recurrent neural network. See Using Recurrent Neural Networks.	PPO, SAC, BC

*PPO = Proximal Policy Optimization, SAC = Soft Actor-Critic, BC = Behavioral Cloning (Imitation)

For specific advice on setting hyperparameters based on the type of training you are conducting, see:

• Training with PPO
• Training with SAC
• Using Recurrent Neural Networks
• Training with Curriculum Learning
• Training with Imitation Learning
• Training Generalized Reinforcement Learning Agents

You can also compare the example environments to the corresponding sections of the config/trainer_config.yaml file for each example to see how the hyperparameters and other configuration variables have been changed from the defaults.

Debugging and Profiling

If you enable the --debug flag in the command line, the trainer metrics are logged to a CSV file stored in the summaries directory. The metrics stored are:

• brain name
• time to update policy
• time since start of training
• time for last experience collection
• number of experiences used for training
• mean return

This option is not available currently for Behavioral Cloning.

Additionally, we have included basic Profiling in Python as part of the toolkit. This information is also saved in the summaries directory.