Main Content

rlPGAgent

Policy gradient reinforcement learning agent

Description

The policy gradient (PG) algorithm is a model-free, online, on-policy reinforcement learning method. A PG agent is a policy-based reinforcement learning agent that uses the REINFORCE algorithm to directly compute an optimal policy which maximizes the long-term reward. The action space can be either discrete or continuous.

For more information on PG agents and the REINFORCE algorithm, see Policy Gradient Agents. For more information on the different types of reinforcement learning agents, see Reinforcement Learning Agents.

Creation

Description

Create Agent from Observation and Action Specifications

example

agent = rlPGAgent(observationInfo,actionInfo) creates a policy gradient agent for an environment with the given observation and action specifications, using default initialization options. The actor and critic in the agent use default deep neural networks built from the observation specification observationInfo and the action specification actionInfo. The ObservationInfo and ActionInfo properties of agent are set to the observationInfo and actionInfo input arguments, respectively.

example

agent = rlPGAgent(observationInfo,actionInfo,initOpts) creates a policy gradient agent for an environment with the given observation and action specifications. The agent uses default networks in which each hidden fully connected layer has the number of units specified in the initOpts object. Policy gradient agents do not support recurrent neural networks. For more information on the initialization options, see rlAgentInitializationOptions.

Create Agent from Actor and Critic

agent = rlPGAgent(actor) creates a PG agent with the specified actor network. By default, the UseBaseline property of the agent is false in this case.

agent = rlPGAgent(actor,critic) creates a PG agent with the specified actor and critic networks. By default, the UseBaseline option is true in this case.

Specify Agent Options

example

agent = rlPGAgent(___,agentOptions) creates a PG agent and sets the AgentOptions property to the agentOptions input argument. Use this syntax after any of the input arguments in the previous syntaxes.

Input Arguments

expand all

Agent initialization options, specified as an rlAgentInitializationOptions object. Policy gradient agents do not support recurrent neural networks.

Actor that implements the policy, specified as an rlDiscreteCategoricalActor or rlContinuousGaussianActor function approximator object. For more information on creating actor approximators, see Create Policies and Value Functions.

Baseline critic that estimates the discounted long-term reward, specified as an rlValueFunction object. For more information on creating critic approximators, see Create Policies and Value Functions.

Properties

expand all

Observation specifications, specified as a reinforcement learning specification object or an array of specification objects defining properties such as dimensions, data type, and names of the observation signals.

If you create the agent by specifying an actor and critic, the value of ObservationInfo matches the value specified in the actor and critic objects.

You can extract observationInfo from an existing environment or agent using getObservationInfo. You can also construct the specifications manually using rlFiniteSetSpec or rlNumericSpec.

Action specifications, specified as a reinforcement learning specification object defining properties such as dimensions, data type, and names of the action signals.

For a discrete action space, you must specify actionInfo as an rlFiniteSetSpec object.

For a continuous action space, you must specify actionInfo as an rlNumericSpec object.

If you create the agent by specifying an actor and critic, the value of ActionInfo matches the value specified in the actor and critic objects.

You can extract actionInfo from an existing environment or agent using getActionInfo. You can also construct the specification manually using rlFiniteSetSpec or rlNumericSpec.

Agent options, specified as an rlPGAgentOptions object.

Option to use exploration policy when selecting actions, specified as a one of the following logical values.

  • true — Use the base agent exploration policy when selecting actions.

  • false — Use the base agent greedy policy when selecting actions.

Sample time of agent, specified as a positive scalar or as -1. Setting this parameter to -1 allows for event-based simulations. The value of SampleTime matches the value specified in AgentOptions.

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.

Object Functions

trainTrain reinforcement learning agents within a specified environment
simSimulate trained reinforcement learning agents within specified environment
getActionObtain action from agent or actor given environment observations
getActorGet actor from reinforcement learning agent
setActorSet actor of reinforcement learning agent
getCriticGet critic from reinforcement learning agent
setCriticSet critic of reinforcement learning agent
generatePolicyFunctionCreate function that evaluates trained policy of reinforcement learning agent

Examples

collapse all

Create an environment with a discrete action space, and obtain its observation and action specifications. For this example, load the environment used in the example Create Agent Using Deep Network Designer and Train Using Image Observations. This environment has two observations: a 50-by-50 grayscale image and a scalar (the angular velocity of the pendulum). The action is a scalar with five possible elements (a torque of either -2, -1, 0, 1, or 2 Nm applied to the pole).

% load predefined environment
env = rlPredefinedEnv("SimplePendulumWithImage-Discrete");

% obtain observation and action specifications
obsInfo = getObservationInfo(env);
actInfo = getActionInfo(env);

The agent creation function initializes the actor and critic networks randomly. You can ensure reproducibility by fixing the seed of the random generator. To do so, uncomment the following line.

% rng(0)

Create a policy gradient agent from the environment observation and action specifications.

agent = rlPGAgent(obsInfo,actInfo);

To check your agent, use getAction to return the action from a random observation.

getAction(agent,{rand(obsInfo(1).Dimension),rand(obsInfo(2).Dimension)})
ans = 1x1 cell array
    {[-2]}

You can now test and train the agent within the environment.

Create an environment with a continuous action space and obtain its observation and action specifications. For this example, load the environment used in the example Train DDPG Agent to Swing Up and Balance Pendulum with Image Observation. This environment has two observations: a 50-by-50 grayscale image and a scalar (the angular velocity of the pendulum). The action is a scalar representing a torque ranging continuously from -2 to 2 Nm.

% load predefined environment
env = rlPredefinedEnv("SimplePendulumWithImage-Continuous");

% obtain observation and action specifications
obsInfo = getObservationInfo(env);
actInfo = getActionInfo(env);

Create an agent initialization option object, specifying that each hidden fully connected layer in the network must have 128 neurons (instead of the default number, 256). Policy gradient agents do not support recurrent networks, so setting the UseRNN option to true generates an error when the agent is created.

initOpts = rlAgentInitializationOptions('NumHiddenUnit',128);

The agent creation function initializes the actor and critic networks randomly. You can ensure reproducibility by fixing the seed of the random generator. To do so, uncomment the following line.

% rng(0)

Create a policy gradient agent from the environment observation and action specifications.

agent = rlPGAgent(obsInfo,actInfo,initOpts);

Extract the deep neural networks from both the agent actor and critic.

actorNet = getModel(getActor(agent));
criticNet = getModel(getCritic(agent));

Display the layers of the critic network, and verify that each hidden fully connected layer has 128 neurons

criticNet.Layers
ans = 
  11x1 Layer array with layers:

     1   'concat'         Concatenation     Concatenation of 2 inputs along dimension 1
     2   'relu_body'      ReLU              ReLU
     3   'fc_body'        Fully Connected   128 fully connected layer
     4   'body_output'    ReLU              ReLU
     5   'input_1'        Image Input       50x50x1 images
     6   'conv_1'         Convolution       64 3x3x1 convolutions with stride [1  1] and padding [0  0  0  0]
     7   'relu_input_1'   ReLU              ReLU
     8   'fc_1'           Fully Connected   128 fully connected layer
     9   'input_2'        Feature Input     1 features
    10   'fc_2'           Fully Connected   128 fully connected layer
    11   'output'         Fully Connected   1 fully connected layer

Plot actor and critic networks

plot(layerGraph(actorNet))

Figure contains an axes object. The axes object contains an object of type graphplot.

plot(layerGraph(criticNet))

Figure contains an axes object. The axes object contains an object of type graphplot.

To check your agent, use getAction to return the action from a random observation.

getAction(agent,{rand(obsInfo(1).Dimension),rand(obsInfo(2).Dimension)})
ans = 1x1 cell array
    {[0.9228]}

You can now test and train the agent within the environment.

Create an environment with a discrete action space, and obtain its observation and action specifications. For this example, load the environment used in the example Train PG Agent with Baseline to Control Double Integrator System. The observation from the environment is a vector containing the position and velocity of a mass. The action is a scalar representing a force, applied to the mass, having three possible values (-2, 0, or 2 Newton).

env = rlPredefinedEnv("DoubleIntegrator-Discrete");
obsInfo = getObservationInfo(env);
actInfo = getActionInfo(env);

Create a neural network to be used as approximation model within the baseline critic. For policy gradient agents, the baseline critic estimates a value function, therefore it must take the observation signal as input and return a scalar value.

baselineNetwork = [
    featureInputLayer(prod(obsInfo.Dimension), ...
        'Normalization', 'none', 'Name', 'state')
    fullyConnectedLayer(8, 'Name', 'BaselineFC')
    reluLayer('Name', 'CriticRelu1')
    fullyConnectedLayer(1, 'Name', 'BaselineFC2', ...
        'BiasLearnRateFactor', 0)];

Create a critic to use as a baseline. Policy gradient agents use an rlValueFunction object to implement the critic.

baseline = rlValueFunction(baselineNetwork,obsInfo);

Set some training options for the critic.

baselineOpts = rlOptimizerOptions( ...
    'LearnRate',5e-3,'GradientThreshold',1);

Create a deep neural network to be used as approximation model within the actor. For policy gradient agents, the actor executes a stochastic policy, which for discrete action spaces is implemented by a discrete categorical actor. In this case the network must take the observation signal as input and return a probability for each action. Therefore the output layer must have as many elements as the number of possible actions.

actorNetwork = [
    featureInputLayer(prod(obsInfo.Dimension), ...
        'Normalization', 'none', 'Name', 'state')
    fullyConnectedLayer(numel(actInfo.Elements), ...
        'Name', 'action', 'BiasLearnRateFactor', 0)];

Create the actor using rlDiscreteCategoricalActor, as well as the observation and action specifications.

actor = rlDiscreteCategoricalActor(actorNetwork,obsInfo,actInfo);

Set some training options for the actor.

actorOpts = rlOptimizerOptions( ...
    'LearnRate',5e-3,'GradientThreshold',1);

Specify agent options.

agentOpts = rlPGAgentOptions( ...
    'UseBaseline',true, ...
    'DiscountFactor', 0.99, ...
    'CriticOptimizerOptions',baselineOpts, ...
    'ActorOptimizerOptions',actorOpts);

Create a PG agent using the actor, critic, and the agent options object.

agent = rlPGAgent(actor,baseline,agentOpts)
agent = 
  rlPGAgent with properties:

            AgentOptions: [1x1 rl.option.rlPGAgentOptions]
    UseExplorationPolicy: 1
         ObservationInfo: [1x1 rl.util.rlNumericSpec]
              ActionInfo: [1x1 rl.util.rlFiniteSetSpec]
              SampleTime: 1

To check your agent, use getAction to return the action from a random observation.

getAction(agent,{rand(obsInfo.Dimension)})
ans = 1x1 cell array
    {[-2]}

You can now test and train the agent within the environment.

Create an environment with a continuous action space, and obtain its observation and action specifications. For this example, load the double integrator continuous action space environment used in the example Train DDPG Agent to Control Double Integrator System.

env = rlPredefinedEnv("DoubleIntegrator-Continuous");
obsInfo = getObservationInfo(env)
obsInfo = 
  rlNumericSpec with properties:

     LowerLimit: -Inf
     UpperLimit: Inf
           Name: "states"
    Description: "x, dx"
      Dimension: [2 1]
       DataType: "double"

actInfo = getActionInfo(env)
actInfo = 
  rlNumericSpec with properties:

     LowerLimit: -Inf
     UpperLimit: Inf
           Name: "force"
    Description: [0x0 string]
      Dimension: [1 1]
       DataType: "double"

In this example, the action is a scalar input representing a force ranging from -2 to 2 Newton, so it is a good idea to set the upper and lower limit of the action signal accordingly. This must be done when the network for the actor has an nonlinear output layer than needs to be scaled accordingly to produce an output in the desired range.

actInfo.LowerLimit=-2;
actInfo.UpperLimit=2;

Policy gradient agents use a value function critic for the baseline. For policy gradient agents, the baseline critic estimates a value function, therefore it must take the observation signal as input and return a scalar value.

baselineNetwork = [
    featureInputLayer(prod(obsInfo.Dimension), ...
        'Normalization', 'none', 'Name', 'state')
    fullyConnectedLayer(8, 'Name', 'bFC1')
    reluLayer('Name', 'bRelu1')
    fullyConnectedLayer(1, 'Name', 'bFC2', ...
        'BiasLearnRateFactor', 0)];

Create a critic to use as a baseline. Policy gradient agents use an rlValueFunction object to implement the critic.

baseline = rlValueFunction(baselineNetwork,obsInfo);

Set some training options for the critic.

baselineOpts = rlOptimizerOptions( ...
    'LearnRate',5e-3,'GradientThreshold',1);

Create a deep neural network to be used as approximation model within the actor. For policy gradient agents, the actor executes a stochastic policy, which for continuous action spaces is implemented by a continuous Gaussian actor. In this case the network must take the observation signal as input and return both a mean value and a standard deviation value for each action. Therefore it must have two output layers (one for the mean values the other for the standard deviation values), each having as many elements as the dimension of the action space.

Note that standard deviations must be nonnegative and mean values must fall within the range of the action. Therefore the output layer that returns the standard deviations must be a softplus or ReLU layer, to enforce nonnegativity, while the output layer that returns the mean values must be a scaling layer, to scale the mean values to the output range.

% input path layers (1 by 1 output)
inPath = [ 
    featureInputLayer(prod(obsInfo.Dimension), ...
        'Normalization','none','Name','state')
    fullyConnectedLayer(10,'Name', 'ip_fc')   
    reluLayer('Name', 'ip_relu')          
    fullyConnectedLayer(1,'Name','ip_out') ];

% mean path layers (1 by 1 output)
meanPath = [
    fullyConnectedLayer(15,'Name', 'mp_fc1')
    reluLayer('Name', 'mp_relu')             
    fullyConnectedLayer(1,'Name','mp_fc2');
    tanhLayer('Name','tanh'); % out range: -1,1
    scalingLayer('Name','mp_out', ...
        'Scale',actInfo.UpperLimit) ]; % range: -2,2

% std path layers (1 by 1 output)
sdevPath = [
    fullyConnectedLayer(15,'Name', 'vp_fc1')
    reluLayer('Name', 'vp_relu')            
    fullyConnectedLayer(1,'Name','vp_fc2'); 
    softplusLayer('Name', 'vp_out') ]; % non negative

% add layers to layerGraph network object
actorNet = layerGraph(inPath);
actorNet = addLayers(actorNet,meanPath);
actorNet = addLayers(actorNet,sdevPath);

% connect output of inPath to meanPath input
actorNet = connectLayers(actorNet,'ip_out','mp_fc1/in');
% connect output of inPath to variancePath input
actorNet = connectLayers(actorNet,'ip_out','vp_fc1/in');

% plot network 
plot(actorNet)

Figure contains an axes object. The axes object contains an object of type graphplot.

Create the actor using rlContinuousGaussianActor, together with actorNet, the observation and action specifications, as well as the names of the network input and output layers.

actor = rlContinuousGaussianActor(actorNet, ...
    obsInfo,actInfo, ...
    'ObservationInputNames',{'state'}, ...
    'ActionMeanOutputNames',{'mp_out'}, ...
    'ActionStandardDeviationOutputNames',{'vp_out'});

Set some training options for the actor.

actorOpts = rlOptimizerOptions( ...
    'LearnRate',5e-3,'GradientThreshold',1);

Specify agent options, and create a PG agent using actor, baseline and agent options.

agentOpts = rlPGAgentOptions(...
    'UseBaseline',true, ...
    'DiscountFactor', 0.99, ...
    'ActorOptimizerOptions', actorOpts);
agent = rlPGAgent(actor,baseline,agentOpts)
agent = 
  rlPGAgent with properties:

            AgentOptions: [1x1 rl.option.rlPGAgentOptions]
    UseExplorationPolicy: 1
         ObservationInfo: [1x1 rl.util.rlNumericSpec]
              ActionInfo: [1x1 rl.util.rlNumericSpec]
              SampleTime: 1

To check your agent, use getAction to return the action from a random observation.

getAction(agent,{rand(obsInfo.Dimension)})
ans = 1x1 cell array
    {[0.0347]}

You can now test and train the agent within the environment.

For this example, load the environment used in the example Train PG Agent with Baseline to Control Double Integrator System. The observation from the environment is a vector containing the position and velocity of a mass. The action is a scalar representing a force, applied to the mass, having three possible values (-2, 0, or 2 Newton).

env = rlPredefinedEnv("DoubleIntegrator-Discrete");

Get observation and specification information.

obsInfo = getObservationInfo(env);
actInfo = getActionInfo(env);

Create a critic to use as a baseline. For policy gradient agents, the baseline critic estimates a value function, therefore it must take the observation signal as input and return a scalar value. To create a recurrent neural network for the critic, use sequenceInputLayer as the input layer and include an lstmLayer as one of the other network layers.

baselineNetwork = [
    sequenceInputLayer(prod(obsInfo.Dimension), ...
        'Normalization', 'none', 'Name','myobs')
    fullyConnectedLayer(8, 'Name', 'BaselineFC')
    lstmLayer(8,'OutputMode','sequence','Name','lstm')
    reluLayer('Name', 'CriticRelu1')
    fullyConnectedLayer(1, 'Name', 'BaselineFC2', ...
        'BiasLearnRateFactor', 0)];

Create the critic based on the network approximator model. Policy gradient agents use an rlValueFunction object to implement the critic.

baseline = rlValueFunction(baselineNetwork,obsInfo);

Set some training options for the critic.

baselineOpts = rlOptimizerOptions( ...
    'LearnRate',5e-3,'GradientThreshold',1);

Since the critic has a recurrent network, the actor must have a recurrent network too. Define a recurrent neural network for the actor. For policy gradient agents, the actor executes a stochastic policy, which for discrete action spaces is implemented by a discrete categorical actor. In this case the network must take the observation signal as input and return a probability for each action. Therefore the output layer must have as many elements as the number of possible actions.

actorNetwork = [
    sequenceInputLayer(prod(obsInfo.Dimension), ...
    'Normalization', 'none', 'Name', 'myobs')
    lstmLayer(8,'OutputMode','sequence','Name','lstm')
    fullyConnectedLayer(numel(actInfo.Elements), ...
    'Name', 'action', 'BiasLearnRateFactor', 0)];

Create the actor. Policy gradient agents use stochastic actors, which for discrete action spaces are implemented by rlDiscreteCategoricalActor objects.

actor = rlDiscreteCategoricalActor(actorNetwork, ...
    obsInfo,actInfo,...
    'Observation',{'myobs'});

Set some training options for the actor.

actorOpts = rlOptimizerOptions( ...
    'LearnRate',5e-3,'GradientThreshold',1);

Specify agent options, and create a PG agent using the actor and critic.

agentOpts = rlPGAgentOptions(...
    'UseBaseline',true, ...
    'DiscountFactor', 0.99, ...
    'CriticOptimizerOptions',baselineOpts, ...
    'ActorOptimizerOptions', actorOpts);
agent = rlPGAgent(actor,baseline);

For PG agent with recurrent neural networks, the training sequence length is the whole episode.

To check your agent, use getAction to return the action from a random observation.

getAction(agent,{rand(obsInfo.Dimension)})
ans = 1x1 cell array
    {[-2]}

You can now test and train the agent within the environment.

Tips

  • For continuous action spaces, the rlPGAgent agent does not enforce the constraints set by the action specification, so you must enforce action space constraints within the environment.

Version History

Introduced in R2019a