Main Content

rlSACAgentOptions

Options for SAC agent

Since R2020b

Description

Use an rlSACAgentOptions object to specify options for soft actor-critic (SAC) agents. To create a SAC agent, use rlSACAgent.

For more information, see Soft Actor-Critic (SAC) Agents.

For more information on the different types of reinforcement learning agents, see Reinforcement Learning Agents.

Creation

Description

opt = rlSACAgentOptions creates an options object for use as an argument when creating a SAC agent using all default options. You can modify the object properties using dot notation.

example

opt = rlSACAgentOptions(Name=Value) creates the options set opt and sets its properties using one or more name-value arguments. For example, rlSACAgentOptions(DiscountFactor=0.95) creates an option set with a discount factor of 0.95. You can specify multiple name-value arguments.

Properties

expand all

Sample time of agent, specified as a positive scalar or as -1. Setting this parameter to -1 allows for event-based simulations.

Within a Simulink® environment, the RL Agent block in which the agent is specified to execute every SampleTime seconds of simulation time. If SampleTime is -1, the block inherits the sample time from its parent subsystem.

Within a MATLAB® environment, the agent is executed every time the environment advances. In this case, SampleTime is the time interval between consecutive elements in the output experience returned by sim or train. If SampleTime is -1, the time interval between consecutive elements in the returned output experience reflects the timing of the event that triggers the agent execution.

This property is shared between the agent and the agent options object within the agent. Therefore, if you change it in the agent options object, it gets changed in the agent, and vice versa.

Example: SampleTime=-1

Discount factor applied to future rewards during training, specified as a positive scalar less than or equal to 1.

Example: DiscountFactor=0.9

Entropy tuning options, specified as an EntropyWeightOptions object with the following properties.

Initial entropy component weight, specified as a positive scalar.

Optimizer learning rate, specified as a nonnegative scalar. If LearnRate is zero, the EntropyWeight value is fixed during training and the TargetEntropy value is ignored.

Target entropy value for tuning entropy weight, specified as a scalar. A higher target entropy value encourages more exploration.

If you do not specify TargetEntropy, the agent uses –A as the target value, where A is the number of actions.

Algorithm to tune entropy, specified as one of the following strings.

  • "adam" — Use the Adam optimizer. You can specify the decay rates of the gradient and squared gradient moving averages using the GradientDecayFactor and SquaredGradientDecayFactor fields of the OptimizerParameters option.

  • "sgdm" — Use the stochastic gradient descent with momentum (SGDM) optimizer. You can specify the momentum value using the Momentum field of the OptimizerParameters option.

  • "rmsprop" — Use the RMSProp optimizer. You can specify the decay rate of the squared gradient moving average using the SquaredGradientDecayFactor fields of the OptimizerParameters option.

For more information about these optimizers, see the Algorithms section of trainingOptions in Deep Learning Toolbox™.

Threshold value for the entropy gradient, specified as Inf or a positive scalar. If the gradient exceeds this value, the gradient is clipped.

Applicable parameters for the optimizer, specified as an OptimizerParameters object with the following parameters. The default parameter values work well for most problems.

ParameterDescriptionDefault
Momentum

Contribution of previous step, specified as a scalar from 0 to 1. A value of 0 means no contribution from the previous step. A value of 1 means maximal contribution.

This parameter applies only when Optimizer is "sgdm".

0.9
Epsilon

Denominator offset, specified as a positive scalar. The optimizer adds this offset to the denominator in the network parameter updates to avoid division by zero.

This parameter applies only when Optimizer is "adam" or "rmsprop".

1e-8
GradientDecayFactor

Decay rate of gradient moving average, specified as a positive scalar from 0 to 1.

This parameter applies only when Optimizer is "adam".

0.9
SquaredGradientDecayFactor

Decay rate of squared gradient moving average, specified as a positive scalar from 0 to 1.

This parameter applies only when Optimizer is "adam" or "rmsprop".

0.999

When a particular property of OptimizerParameters is not applicable to the optimizer type specified in the Algorithm option, that property is set to "Not applicable".

To change the default values, access the properties of OptimizerParameters using dot notation.

opt = rlSACAgentOptions;
opt.EntropyWeightOptions.OptimizerParameters.GradientDecayFactor = 0.95;

Experience buffer size, specified as a positive integer. During training, the agent computes updates using a mini-batch of experiences randomly sampled from the buffer.

Example: ExperienceBufferLength=1e6

Size of random experience mini-batch, specified as a positive integer. During each training episode, the agent randomly samples experiences from the experience buffer when computing gradients for updating the actor and critics. Large mini-batches reduce the variance when computing gradients but increase the computational effort.

Example: MiniBatchSize=128

Maximum batch-training trajectory length when using a recurrent neural network, specified as a positive integer. This value must be greater than 1 when using a recurrent neural network and 1 otherwise.

Example: SequenceLength=4

Actor optimizer options, specified as an rlOptimizerOptions object. It allows you to specify training parameters of the actor approximator such as learning rate, gradient threshold, as well as the optimizer algorithm and its parameters. For more information, see rlOptimizerOptions and rlOptimizer.

Example: ActorOptimizerOptions = rlOptimizerOptions(LearnRate=2e-3)

Critic optimizer options, specified as an rlOptimizerOptions object. It allows you to specify training parameters of the critic approximator such as learning rate, gradient threshold, as well as the optimizer algorithm and its parameters. For more information, see rlOptimizerOptions and rlOptimizer.

Example: CriticOptimizerOptions = rlOptimizerOptions(LearnRate=5e-3)

Number of future rewards used to estimate the value of the policy, specified as a positive integer. Specifically, ifNumStepsToLookAhead is equal to N, the target value of the policy at a given step is calculated adding the rewards for the following N steps and the discounted estimated value of the state that caused the N-th reward. This target is also called N-step return.

Note

When using a recurrent neural network for the critic, NumStepsToLookAhead must be 1.

For more information, see [1], Chapter 7.

Example: NumStepsToLookAhead=3

Minimum number of samples to generate before learning starts. Use this option to ensure that learning takes place over a more diverse data set at the beginning of training. The default, and minimum, value is the value of MiniBatchSize. After the software collects a minimum of NumWarmStartSteps samples, learning occurs at the intervals specified by the LearningFrequency property.

Example: NumWarmStartSteps=20

Number of times an agent learns over a data set, specified as a positive integer. For off-policy agents that support this property (DQN, DDPG, TD3 and SAC), this value defines the number of passes over the data in the replay buffer at each learning iteration.

Example: NumEpoch=2

Maximum number of mini-batches used for learning during a single epoch, specified as a positive integer.

For off-policy agents that support this property (DQN, DDPG, TD3, and SAC), the actual number of mini-batches used for learning depends on the length of the replay buffer, and MaxMiniBatchPerEpoch specifies the upper bound. This value also specifies the maximum number of gradient steps per learning iteration because the maximum number of gradient steps is equal to the MaxMiniBatchPerEpoch value multiplied by the NumEpoch value.

For off-policy agents that support this property, a high MaxMiniBatchPerEpoch value means that more time is spent on learning than collecting new data. Therefore, you can use this parameter to control the sample efficiency of the learning process.

Example: MaxMiniBatchPerEpoch=200

Minimum number of environment interactions between learning iterations, specified as a positive integer or -1. This value defines how many new data samples need to be generated before learning. For off-policy agents that support this property (DQN, DDPG, TD3, and SAC), the default value of -1 means that learning occurs after each episode is finished. Note that learning can start only after the software collects a minimum of NumWarmStartSteps samples. It then occurs at the intervals specified by the LearningFrequency property.

Example: LearningFrequency=4

Period of policy update with respect to critic update, specified as a positive integer. This option defines how often the actor is updated with respect to each critic update. For example, a value of 3 means that the actor is updated every three critic updates. Updating the actor less frequently than the critic can improve convergence at the cost of longer training times.

Example: PolicyUpdateFrequency=2

Number of steps between target critic updates, specified as a positive integer. For more information, see Target Update Methods.

Example: TargetUpdateFrequency=5

Smoothing factor for target critic updates, specified as a positive scalar less than or equal to 1. For more information, see Target Update Methods.

Example: TargetSmoothFactor=1e-2

Option to use entropy in the critic targets, specified as a either true (default, entropy is used) or false (entropy is not used). Note that this option does not affect entropy usage in the actor.

Example: UseCriticTargetEntropy=false

Batch data regularizer options, specified as an rlBehaviorCloningRegularizerOptions object. These options are typically used to train the agent offline, from existing data. If you leave this option empty, no regularizer is used.

For more information, see rlBehaviorCloningRegularizerOptions.

Example: BatchDataRegularizerOptions = rlBehaviorCloningRegularizerOptions(BehaviorCloningRegularizerWeight=10)

Option for clearing the experience buffer before training, specified as a logical value.

Example: ResetExperienceBufferBeforeTraining=true

Options to save additional agent data, specified as a structure containing the following fields.

  • Optimizer

  • PolicyState

  • Target

  • ExperienceBuffer

You can save an agent object in one of the following ways:

  • Using the save command

  • Specifying saveAgentCriteria and saveAgentValue in an rlTrainingOptions object

  • Specifying an appropriate logging function within a FileLogger object.

When you save an agent using any method, the fields in the InfoToSave structure determine whether the corresponding data is saved with the agent. For example, if you set the Optimizer field to true, then the actor and critic optimizers are saved along with the agent.

You can modify the InfoToSave property only after the agent options object is created.

Example: options.InfoToSave.Optimizer=true

Option to save the actor and critic optimizers, specified as a logical value. If you set the Optimizer field to false, then the actor and critic optimizers (which are hidden properties of the agent and can contain internal states) are not saved along with the agent, therefore saving disk space and memory. However, when the optimizers contains internal states, the state of the saved agent is not identical to the state of the original agent.

Example: true

Option to save the state of the explorative policy, specified as a logical value. If you set the PolicyState field to false, then the state of the explorative policy (which is a hidden agent property) is not saved along with the agent. In this case, the state of the saved agent is not identical to the state of the original agent.

Example: true

Option to save the actor and critic targets, specified as a logical value. If you set the Target field to false, then the actor and critic targets (which are hidden agent properties) is not saved along with the agent. In this case, when the targets contain internal states, the state of the saved agent is not identical to the state of the original agent.

Example: true

Option to save the experience buffer, specified as a logical value. If you set the PolicyState field to false, then the content of the experience buffer (which is accessible as an agent property using dot notation) is not saved along with the agent. In this case, the state of the saved agent is not identical to the state of the original agent.

Example: true

Object Functions

rlSACAgentSoft actor-critic (SAC) reinforcement learning agent

Examples

collapse all

Create a SAC agent options object, specifying the discount factor.

opt = rlSACAgentOptions(DiscountFactor=0.95)
opt = 
  rlSACAgentOptions with properties:

                             SampleTime: 1
                         DiscountFactor: 0.9500
                   EntropyWeightOptions: [1x1 rl.option.EntropyWeightOptions]
                 ExperienceBufferLength: 10000
                          MiniBatchSize: 64
                         SequenceLength: 1
                  ActorOptimizerOptions: [1x1 rl.option.rlOptimizerOptions]
                 CriticOptimizerOptions: [1x2 rl.option.rlOptimizerOptions]
                    NumStepsToLookAhead: 1
                      NumWarmStartSteps: 64
                               NumEpoch: 1
                   MaxMiniBatchPerEpoch: 100
                      LearningFrequency: -1
                  PolicyUpdateFrequency: 1
                  TargetUpdateFrequency: 1
                     TargetSmoothFactor: 1.0000e-03
                 UseCriticTargetEntropy: 1
            BatchDataRegularizerOptions: []
    ResetExperienceBufferBeforeTraining: 0
                             InfoToSave: [1x1 struct]

You can modify options using dot notation. For example, set the agent sample time to 0.5.

opt.SampleTime = 0.5;

For SAC agents, configure the entropy weight optimizer using the options in EntropyWeightOptions. For example, set the target entropy value to –5.

opt.EntropyWeightOptions.TargetEntropy = -5;

References

[1] Sutton, Richard S., and Andrew G. Barto. Reinforcement Learning: An Introduction. Second edition. Adaptive Computation and Machine Learning. Cambridge, Mass: The MIT Press, 2018.

Version History

Introduced in R2020b

expand all