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 use the generated model file with the Internal Brain type in your Unity project to decide the best course of action for an agent.

Use the Python program, learn.py to train your agents. This program can be found in the python directory of the ML-Agents SDK. The configuration file, 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 brain.

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

Training with learn.py

Use the Python learn.py program to train agents. learn.py supports training with reinforcement learning, curriculum learning, and behavioral cloning imitation learning.

Run learn.py from the command line to launch the training process. Use the command line patterns and the trainer_config.yaml file to control training options.

The basic command for training is:

python3 learn.py <env_name> --run-id=<run-identifier> --train

where

• <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 ml-agents python folder.

4. Run the following to launch the training process using the path to the Unity environment you built in step 1:

    python3 learn.py ../../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 (from the ML-Agents python directory):

tensorboard --logdir=summaries

And then opening the URL: localhost:6006.

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

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 learn.py:

• --curriculum=<file> – Specify a curriculum JSON file for defining the lessons for curriculum training. See Curriculum 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.
• --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 python/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.
• --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.
• --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.
• --worker-id=<n> – When you are running more than one training environment at the same time, assign each a unique worker-id number. The worker-id is added to the communication port opened between the current instance of learn.py and the ExternalCommunicator object in the Unity environment. Defaults to 0.
• --docker-target-name=<dt> – The Docker Volume on which to store curriculum, executable and model files. See Using Docker.
• --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.

Training config file

The training config file, trainer_config.yaml specifies the training method, the hyperparameters, and a few additional values to use during training. The file is 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 Brains. Name each of these override sections after the GameObject containing the Brain component that should use these settings. (This GameObject will be a child of the Academy in your scene.) Sections for the example environments are included in the provided config file. Learn.py finds the config file by name and looks for it in the same directory as itself.

** Setting **	Description	Applies To Trainer
batch_size	The number of experiences in each iteration of gradient descent.	PPO, 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, BC
brain_to_imitate	For imitation learning, the name of the GameObject containing the Brain component to imitate.	BC
buffer_size	The number of experiences to collect before updating the policy model.	PPO
curiosity_enc_size	The size of the encoding to use in the forward and inverse models in the Curioity module.	PPO
curiosity_strength	Magnitude of intrinsic reward generated by Intrinsic Curiosity Module.	PPO
epsilon	Influences how rapidly the policy can evolve during training.	PPO, BC
gamma	The reward discount rate for the Generalized Advantage Estimator (GAE).	PPO
hidden_units	The number of units in the hidden layers of the neural network.	PPO, BC
lambd	The regularization parameter.	PPO
learning_rate	The initial learning rate for gradient descent.	PPO, BC
max_steps	The maximum number of simulation steps to run during a training session.	PPO, 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, BC
normalize	Whether to automatically normalize observations.	PPO, BC
num_epoch	The number of passes to make through the experience buffer when performing gradient descent optimization.	PPO, BC
num_layers	The number of hidden layers in the neural network.	PPO, BC
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, BC
summary_freq	How often, in steps, to save training statistics. This determines the number of data points shown by TensorBoard.	PPO, BC
time_horizon	How many steps of experience to collect per-agent before adding it to the experience buffer.	PPO, BC
trainer	The type of training to perform: "ppo" or "imitation".	PPO, BC
use_curiosity	Train using an additional intrinsic reward signal generated from Intrinsic Curiosity Module.	PPO
use_recurrent	Train using a recurrent neural network. See Using Recurrent Neural Networks.	PPO, BC
	PPO = Proximal Policy Optimization, BC = Behavioral Cloning (Imitation))

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

• Training with PPO
• Using Recurrent Neural Networks
• Training with Curriculum Learning
• Training with Imitation Learning

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