ml-agents/docs/Training-PPO.md

Training with Proximal Policy Optimization

ML-Agents provides an implementation of a reinforcement learning algorithm called Proximal Policy Optimization (PPO). PPO uses a neural network to approximate the ideal function that maps an agent's observations to the best action an agent can take in a given state. The ML-Agents PPO algorithm is implemented in TensorFlow and runs in a separate Python process (communicating with the running Unity application over a socket).

ML-Agents also provides an implementation of Soft Actor-Critic (SAC). SAC tends to be more sample-efficient, i.e. require fewer environment steps, than PPO, but may spend more time performing model updates. This can produce a large speedup on heavy or slow environments. Check out how to train with SAC here.

To train an agent, you will need to provide the agent one or more reward signals which the agent should attempt to maximize. See Reward Signals for the available reward signals and the corresponding hyperparameters.

See Training ML-Agents for instructions on running the training program, learn.py.

If you are using the recurrent neural network (RNN) to utilize memory, see Using Recurrent Neural Networks for RNN-specific training details.

If you are using curriculum training to pace the difficulty of the learning task presented to an agent, see Training with Curriculum Learning.

For information about imitation learning from demonstrations, see Training with Imitation Learning.

Best Practices Training with PPO

Successfully training a Reinforcement Learning model often involves tuning the training hyperparameters. This guide contains some best practices for tuning the training process when the default parameters don't seem to be giving the level of performance you would like.

Hyperparameters

Reward Signals

In reinforcement learning, the goal is to learn a Policy that maximizes reward. At a base level, the reward is given by the environment. However, we could imagine rewarding the agent for various different behaviors. For instance, we could reward the agent for exploring new states, rather than just when an explicit reward is given. Furthermore, we could mix reward signals to help the learning process.

Using reward_signals allows you to define reward signals. The ML-Agents Toolkit provides three reward signals by default, the Extrinsic (environment) reward signal, the Curiosity reward signal, which can be used to encourage exploration in sparse extrinsic reward environments, and the GAIL reward signal. Please see Reward Signals for additional details.

Lambda

lambd corresponds to the lambda parameter used when calculating the Generalized Advantage Estimate (GAE). This can be thought of as how much the agent relies on its current value estimate when calculating an updated value estimate. Low values correspond to relying more on the current value estimate (which can be high bias), and high values correspond to relying more on the actual rewards received in the environment (which can be high variance). The parameter provides a trade-off between the two, and the right value can lead to a more stable training process.

Typical Range: 0.9 - 0.95

Buffer Size

buffer_size corresponds to how many experiences (agent observations, actions and rewards obtained) should be collected before we do any learning or updating of the model. This should be a multiple of batch_size. Typically a larger buffer_size corresponds to more stable training updates.

Typical Range: 2048 - 409600

Batch Size

batch_size is the number of experiences used for one iteration of a gradient descent update. This should always be a fraction of the buffer_size. If you are using a continuous action space, this value should be large (in the order of 1000s). If you are using a discrete action space, this value should be smaller (in order of 10s).

Typical Range (Continuous): 512 - 5120

Typical Range (Discrete): 32 - 512

Number of Epochs

num_epoch is the number of passes through the experience buffer during gradient descent. The larger the batch_size, the larger it is acceptable to make this. Decreasing this will ensure more stable updates, at the cost of slower learning.

Typical Range: 3 - 10

Learning Rate

learning_rate corresponds to the strength of each gradient descent update step. This should typically be decreased if training is unstable, and the reward does not consistently increase.

Typical Range: 1e-5 - 1e-3

(Optional) Learning Rate Schedule

learning_rate_schedule corresponds to how the learning rate is changed over time. For PPO, we recommend decaying learning rate until max_steps so learning converges more stably. However, for some cases (e.g. training for an unknown amount of time) this feature can be disabled.

Options:

• linear (default): Decay learning_rate linearly, reaching 0 at max_steps.
• constant: Keep learning rate constant for the entire training run.

Options: linear, constant

Time Horizon

time_horizon corresponds to how many steps of experience to collect per-agent before adding it to the experience buffer. When this limit is reached before the end of an episode, a value estimate is used to predict the overall expected reward from the agent's current state. As such, this parameter trades off between a less biased, but higher variance estimate (long time horizon) and more biased, but less varied estimate (short time horizon). In cases where there are frequent rewards within an episode, or episodes are prohibitively large, a smaller number can be more ideal. This number should be large enough to capture all the important behavior within a sequence of an agent's actions.

Typical Range: 32 - 2048

Max Steps

max_steps corresponds to how many steps of the simulation (multiplied by frame-skip) are run during the training process. This value should be increased for more complex problems.

Typical Range: 5e5 - 1e7

Beta

beta corresponds to the strength of the entropy regularization, which makes the policy "more random." This ensures that agents properly explore the action space during training. Increasing this will ensure more random actions are taken. This should be adjusted such that the entropy (measurable from TensorBoard) slowly decreases alongside increases in reward. If entropy drops too quickly, increase beta. If entropy drops too slowly, decrease beta.

Typical Range: 1e-4 - 1e-2

Epsilon

epsilon corresponds to the acceptable threshold of divergence between the old and new policies during gradient descent updating. Setting this value small will result in more stable updates, but will also slow the training process.

Typical Range: 0.1 - 0.3

Normalize

normalize corresponds to whether normalization is applied to the vector observation inputs. This normalization is based on the running average and variance of the vector observation. Normalization can be helpful in cases with complex continuous control problems, but may be harmful with simpler discrete control problems.

Number of Layers

num_layers corresponds to how many hidden layers are present after the observation input, or after the CNN encoding of the visual observation. For simple problems, fewer layers are likely to train faster and more efficiently. More layers may be necessary for more complex control problems.

Typical range: 1 - 3

Hidden Units

hidden_units correspond to how many units are in each fully connected layer of the neural network. For simple problems where the correct action is a straightforward combination of the observation inputs, this should be small. For problems where the action is a very complex interaction between the observation variables, this should be larger.

Typical Range: 32 - 512

(Optional) Visual Encoder Type

vis_encode_type corresponds to the encoder type for encoding visual observations. Valid options include:

• simple (default): a simple encoder which consists of two convolutional layers
• nature_cnn: CNN implementation proposed by Mnih et al., consisting of three convolutional layers
• resnet: IMPALA Resnet implementation, consisting of three stacked layers, each with two residual blocks, making a much larger network than the other two.

Options: simple, nature_cnn, resnet

(Optional) Recurrent Neural Network Hyperparameters

The below hyperparameters are only used when use_recurrent is set to true.

Sequence Length

sequence_length corresponds to the length of the sequences of experience passed through the network during training. This should be long enough to capture whatever information your agent might need to remember over time. For example, if your agent needs to remember the velocity of objects, then this can be a small value. If your agent needs to remember a piece of information given only once at the beginning of an episode, then this should be a larger value.

Typical Range: 4 - 128

Memory Size

memory_size corresponds to the size of the array of floating point numbers used to store the hidden state of the recurrent neural network of the policy. This value must be a multiple of 2, and should scale with the amount of information you expect the agent will need to remember in order to successfully complete the task.

Typical Range: 32 - 256

(Optional) Behavioral Cloning Using Demonstrations

In some cases, you might want to bootstrap the agent's policy using behavior recorded from a player. This can help guide the agent towards the reward. Behavioral Cloning (BC) adds training operations that mimic a demonstration rather than attempting to maximize reward.

To use BC, add a behavioral_cloning section to the trainer_config. For instance:

    behavioral_cloning:
        demo_path: ./Project/Assets/ML-Agents/Examples/Pyramids/Demos/ExpertPyramid.demo
        strength: 0.5
        steps: 10000

Below are the available hyperparameters for BC.

Strength

strength corresponds to the learning rate of the imitation relative to the learning rate of PPO, and roughly corresponds to how strongly we allow BC to influence the policy.

Typical Range: 0.1 - 0.5

Demo Path

demo_path is the path to your .demo file or directory of .demo files. See the imitation learning guide for more on .demo files.

Steps

During BC, it is often desirable to stop using demonstrations after the agent has "seen" rewards, and allow it to optimize past the available demonstrations and/or generalize outside of the provided demonstrations. steps corresponds to the training steps over which BC is active. The learning rate of BC will anneal over the steps. Set the steps to 0 for constant imitation over the entire training run.

(Optional) Batch Size

batch_size is the number of demonstration experiences used for one iteration of a gradient descent update. If not specified, it will default to the batch_size defined for PPO.

Typical Range (Continuous): 512 - 5120

Typical Range (Discrete): 32 - 512

(Optional) Number of Epochs

num_epoch is the number of passes through the experience buffer during gradient descent. If not specified, it will default to the number of epochs set for PPO.

Typical Range: 3 - 10

(Optional) Samples Per Update

samples_per_update is the maximum number of samples to use during each imitation update. You may want to lower this if your demonstration dataset is very large to avoid overfitting the policy on demonstrations. Set to 0 to train over all of the demonstrations at each update step.

Default Value: 0 (all)

Typical Range: Approximately equal to PPO's buffer_size

(Optional) Advanced: Initialize Model Path

init_path can be specified to initialize your model from a previous run before starting. Note that the prior run should have used the same trainer configurations as the current run, and have been saved with the same version of ML-Agents. You should provide the full path to the folder where the checkpoints were saved, e.g. ./models/{run-id}/{behavior_name}.

This option is provided in case you want to initialize different behaviors from different runs; in most cases, it is sufficient to use the --initialize-from CLI parameter to initialize all models from the same run.

(Optional) Advanced: Disable Threading

By default, PPO model updates can happen while the environment is being stepped. This violates the on-policy assumption of PPO slightly in exchange for a 10-20% training speedup. To maintain the strict on-policyness of PPO, you can disable parallel updates by setting threaded to false.

Default Value: true

Training Statistics

To view training statistics, use TensorBoard. For information on launching and using TensorBoard, see here.

Cumulative Reward

The general trend in reward should consistently increase over time. Small ups and downs are to be expected. Depending on the complexity of the task, a significant increase in reward may not present itself until millions of steps into the training process.

Entropy

This corresponds to how random the decisions are. This should consistently decrease during training. If it decreases too soon or not at all, beta should be adjusted (when using discrete action space).

Learning Rate

This will decrease over time on a linear schedule by default, unless learning_rate_schedule is set to constant.

Policy Loss

These values will oscillate during training. Generally they should be less than 1.0.

Value Estimate

These values should increase as the cumulative reward increases. They correspond to how much future reward the agent predicts itself receiving at any given point.

Value Loss

These values will increase as the reward increases, and then should decrease once reward becomes stable.