rlPPOAgent

Proximal policy optimization (PPO) reinforcement learning agent

Since R2019b

Description

Proximal policy optimization (PPO) is a model-free, online, on-policy, policy gradient reinforcement learning method. This algorithm alternates between sampling data through environmental interaction and optimizing a clipped surrogate objective function using stochastic gradient descent. The action space can be either discrete or continuous. For continuous action spaces, this agent does not enforce constraints set in the action specification; therefore, if you need to enforce action constraints, you must do so within the environment.

For more information on PPO agents, see Proximal Policy Optimization (PPO) Agents. For more information on the different types of reinforcement learning agents, see Reinforcement Learning Agents.

Creation

Syntax

agent = rlPPOAgent(observationInfo,actionInfo)

agent = rlPPOAgent(observationInfo,actionInfo,initOpts)

agent = rlPPOAgent(actor,critic)

agent = rlPPOAgent(___,agentOptions)

Description

Create Agent from Observation and Action Specifications

agent = rlPPOAgent(observationInfo,actionInfo) creates a proximal policy optimization (PPO) 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.

agent = rlPPOAgent(observationInfo,actionInfo,initOpts) creates a PPO agent for an environment with the given observation and action specifications. The agent uses default networks configured using options specified in the initOpts object. Actor-critic agents do not support recurrent neural networks. For more information on the initialization options, see rlAgentInitializationOptions.

Create Agent from Actor and Critic

agent = rlPPOAgent(actor,critic) creates a PPO agent with the specified actor and critic, using the default options for the agent.

Specify Agent Options

agent = rlPPOAgent(___,agentOptions) creates a PPO 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

initOpts — Agent initialization options

rlAgentInitializationOptions object

Agent initialization options, specified as an rlAgentInitializationOptions object.

actor — Actor

rlDiscreteCategoricalActor object | rlContinuousGaussianActor object

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.

critic — Critic

rlValueFunction object

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.

Your critic can use a recurrent neural network as its function approximator. In this case, your actor must also use a recurrent neural network. For an example, see Create PPO Agent with Recurrent Neural Networks.

Properties

ObservationInfo — Observation specifications

specification object | array of specification objects

Observation specifications, specified as an rlFiniteSetSpec or rlNumericSpec object or an array containing a mix of such objects. Each element in the array defines the properties of an environment observation channel, such as its dimensions, data type, and name.

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.

ActionInfo — Action specification

rlFiniteSetSpec object | rlNumericSpec object

Action specifications, specified either as an rlFiniteSetSpec (for discrete action spaces) or rlNumericSpec (for continuous action spaces) object. This object defines the properties of the environment action channel, such as its dimensions, data type, and name.

Note

Only one action channel is allowed.

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.

AgentOptions — Agent options

rlPPOAgentOptions object

Agent options, specified as an rlPPOAgentOptions object.

UseExplorationPolicy — Option to use exploration policy

true (default) | false

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 in sim and generatePolicyFunction. In this case, the agent selects its actions by sampling its probability distribution, the policy is therefore stochastic and the agent explores its observation space.

• false — Use the base agent greedy policy (the action with maximum likelihood) when selecting actions in sim and generatePolicyFunction. In this case, the simulated agent and generated policy behave deterministically.

Note

This option affects only simulation and deployment; it does not affect training.

SampleTime — Sample time of agent

positive scalar | -1

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

train	Train reinforcement learning agents within a specified environment
sim	Simulate trained reinforcement learning agents within specified environment
getAction	Obtain action from agent, actor, or policy object given environment observations
getActor	Extract actor from reinforcement learning agent
setActor	Set actor of reinforcement learning agent
getCritic	Extract critic from reinforcement learning agent
setCritic	Set critic of reinforcement learning agent
generatePolicyFunction	Generate MATLAB function that evaluates policy of an agent or policy object

Examples

Create Discrete PPO Agent from Observation and Action Specifications

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 DQN 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 -2, -1, 0, 1, or 2 Nm applied to a swinging pole).

env = rlPredefinedEnv("SimplePendulumWithImage-Discrete");

Obtain observation and action specifications from the environment.

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

The agent creation function initializes the actor and critic networks randomly. Ensure reproducibility by fixing the seed of the random generator.

rng(0)

Create a PPO agent from the environment observation and action specifications.

agent = rlPPOAgent(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. You can also use getActor and getCritic to extract the actor and critic, respectively, and getModel to extract the approximator model (by default a deep neural network) from the actor or critic.

Create Continuous PPO Agent Using Initialization Options

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.

env = rlPredefinedEnv("SimplePendulumWithImage-Continuous");

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).

initOpts = rlAgentInitializationOptions(NumHiddenUnit=128);

The agent creation function initializes the actor and critic networks randomly. Ensure reproducibility by fixing the seed of the random generator.

rng(0)

Create a PPO actor-critic agent from the environment observation and action specifications.

agent = rlPPOAgent(obsInfo,actInfo,initOpts);

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

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

To verify that each hidden fully connected layer has 128 neurons, you can display the layers on the MATLAB® command window,

criticNet.Layers

or visualize the structure interactively using analyzeNetwork.

analyzeNetwork(criticNet)

Plot actor and critic networks

plot(layerGraph(actorNet))

plot(layerGraph(criticNet))

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 Proximal Policy Optimization Agent

Create an environment interface, and obtain its observation and action specifications.

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

PPO agents use a parametrized value function as a critic. A value-function critic takes the current observation as input and returns a single scalar as output (the estimated discounted cumulative long-term reward for following the policy from the state corresponding to the current observation).

To model the parametrized value function within the critic, use a neural network with one input layer (which receives the content of the observation channel, as specified by obsInfo) and one output layer (which returns the scalar value). Note that prod(obsInfo.Dimension) returns the total number of dimensions of the observation space regardless of whether the observation space is a column vector, row vector, or matrix.

Define the network as an array of layer objects.

criticNet = [
    featureInputLayer(prod(obsInfo.Dimension))
    fullyConnectedLayer(100)
    reluLayer
    fullyConnectedLayer(1)
    ];

Convert to a dlnetwork object and display the number of parameters.

criticNet = dlnetwork(criticNet);
summary(criticNet)

   Initialized: true

   Number of learnables: 601

   Inputs:
      1   'input'   4 features

Create the critic using criticNet and the observation specification object. For more information, see rlValueFunction.

critic = rlValueFunction(criticNet,obsInfo);

Check the critic with a random observation input.

getValue(critic,{rand(obsInfo.Dimension)})

ans = single
    -0.2479

Policy gradient agents use a parametrized stochastic policy, which for discrete action spaces is implemented by a discrete categorical actor. This actor takes an observation as input and returns as output a random action sampled (among the finite number of possible actions) from a categorical probability distribution.

To model the parametrized policy within the actor, use a neural network with one input layer (which receives the content of the environment observation channel, as specified by obsInfo) and one output layer. The output layer must return a vector of probabilities for each possible action, as specified by actInfo. Note that numel(actInfo.Dimension) returns the number of elements of the discrete action space.

Define the network as an array of layer objects.

actorNet = [
    featureInputLayer(prod(obsInfo.Dimension))
    fullyConnectedLayer(200)
    reluLayer
    fullyConnectedLayer(numel(actInfo.Dimension))
    ];

Convert to a dlnetwork object and display the number of parameters.

actorNet = dlnetwork(actorNet);
summary(actorNet)

   Initialized: true

   Number of learnables: 1.4k

   Inputs:
      1   'input'   4 features

Create the actor using actorNet and the environment specification objects. For more information, see rlDiscreteCategoricalActor.

actor = rlDiscreteCategoricalActor(actorNet,obsInfo,actInfo);

Check the actor with a random observation input.

getAction(actor,{rand(obsInfo.Dimension)})

ans = 1x1 cell array
    {[-10]}

Create a PPO agent using the actor and the critic.

agent = rlPPOAgent(actor,critic)

agent = 
  rlPPOAgent with properties:

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

Specify agent options, including training options for the actor and the critic.

agent.AgentOptions.ExperienceHorizon = 1024;
agent.AgentOptions.DiscountFactor = 0.95;

agent.AgentOptions.CriticOptimizerOptions.LearnRate = 8e-3;
agent.AgentOptions.CriticOptimizerOptions.GradientThreshold = 1;
agent.AgentOptions.ActorOptimizerOptions.LearnRate = 8e-3;
agent.AgentOptions.ActorOptimizerOptions.GradientThreshold = 1;

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

getAction(agent,{rand(obsInfo.Dimension)})

ans = 1x1 cell array
    {[-10]}

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

Create Continuous PPO Agent

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. 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, ranging continuously from -2 to 2 Newton.

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 value representing a force ranging from -2 to 2 Newton. To make sure that the output from the agent is in this range, you perform an appropriate scaling operation. Store these limits so you can easily access them later.

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

The actor and critic networks are initialized randomly. Ensure reproducibility by fixing the seed of the random generator.

rng(0)

PPO agents use a parametrized value function as a critic. A value-function critic takes the current observation as input and returns a single scalar as output (the estimated discounted cumulative long-term reward for following the policy from the state corresponding to the current observation).

To model the parametrized value function within the critic, use a neural network with two input layers (one for each observation channel, as specified by obsInfo) and one output layer (returning the scalar value). Note that prod(obsInfo.Dimension) returns the total number of dimensions of the observation space regardless of whether the observation space is a column vector, row vector, or matrix.

Define the network as an array of layer objects.

criticNet = [
    featureInputLayer(prod(obsInfo.Dimension))
    fullyConnectedLayer(100)
    reluLayer
    fullyConnectedLayer(1)
    ];

Convert to a dlnetwork object and display the number of parameters.

criticNet = dlnetwork(criticNet);
summary(criticNet)

   Initialized: true

   Number of learnables: 401

   Inputs:
      1   'input'   2 features

Create the critic approximator object using criticNet and the observation specification. For more information, see rlValueFunction.

critic = rlValueFunction(criticNet,obsInfo);

Check the critic with a random observation input.

getValue(critic,{rand(obsInfo.Dimension)})

ans = single
    -0.0899

PPO agents use a parametrized stochastic policy, which for continuous action spaces is implemented by a continuous Gaussian actor. This actor takes an observation as input and returns as output a random action sampled from a Gaussian probability distribution.

To approximate the mean values and standard deviations of the Gaussian distribution, you must use a neural network with two output layers, each having as many elements as the dimension of the action space. One output layer must return a vector containing the mean values for each action dimension. The other must return a vector containing the standard deviation for each action dimension.

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.

For this example the environment has only one observation channel and therefore the network has only one input layer.

Define each network path as an array of layer objects, and assign names to the input and output layers of each path. These names allow you to connect the paths and then later explicitly associate the network input and output layers with the appropriate environment channel.

% Define common input path layer
commonPath = [ 
    featureInputLayer(prod(obsInfo.Dimension),Name="comPathIn")
    fullyConnectedLayer(100)
    reluLayer
    fullyConnectedLayer(1,Name="comPathOut") 
    ];

% Define mean value path
meanPath = [
    fullyConnectedLayer(15,Name="meanPathIn")
    reluLayer
    fullyConnectedLayer(prod(actInfo.Dimension));
    tanhLayer;
    scalingLayer(Name="meanPathOut",Scale=actInfo.UpperLimit) 
    ];

% Define standard deviation path
sdevPath = [
    fullyConnectedLayer(15,"Name","stdPathIn")
    reluLayer
    fullyConnectedLayer(prod(actInfo.Dimension));
    softplusLayer(Name="stdPathOut") 
    ];

% Add layers to layerGraph object
actorNet = layerGraph(commonPath);
actorNet = addLayers(actorNet,meanPath);
actorNet = addLayers(actorNet,sdevPath);

% Connect paths
actorNet = connectLayers(actorNet,"comPathOut","meanPathIn/in");
actorNet = connectLayers(actorNet,"comPathOut","stdPathIn/in");

% Plot network 
plot(actorNet)

% Convert to dlnetwork and display number of weights
actorNet = dlnetwork(actorNet);
summary(actorNet)

   Initialized: true

   Number of learnables: 493

   Inputs:
      1   'comPathIn'   2 features

Create the actor approximator object using actorNet, the environment specifications, the name of the network input layer to be connected with the environment observation channel, and the names of the network output layers that calculate the mean the standard deviation values of the action.

For more information, see rlContinuousGaussianActor.

actor = rlContinuousGaussianActor(actorNet, obsInfo, actInfo, ...
    "ActionMeanOutputNames","meanPathOut",...
    "ActionStandardDeviationOutputNames","stdPathOut",...
    ObservationInputNames="comPathIn");

Check the actor with a random observation input.

getAction(actor,{rand(obsInfo.Dimension)})

ans = 1x1 cell array
    {[-0.2267]}

Create a PPO agent using the actor and the critic.

agent = rlPPOAgent(actor,critic)

agent = 
  rlPPOAgent with properties:

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

Specify agent options, including training options for the actor and the critic.

agent.AgentOptions.ExperienceHorizon = 1024;
agent.AgentOptions.DiscountFactor = 0.95;

agent.AgentOptions.CriticOptimizerOptions.LearnRate = 8e-3;
agent.AgentOptions.CriticOptimizerOptions.GradientThreshold = 1;
agent.AgentOptions.ActorOptimizerOptions.LearnRate = 8e-3;
agent.AgentOptions.ActorOptimizerOptions.GradientThreshold = 1;

Specify training options for the critic.

criticOpts = rlOptimizerOptions( ...
    LearnRate=8e-3,GradientThreshold=1);

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

getAction(agent,{rand(obsInfo.Dimension)})

ans = 1x1 cell array
    {[0.2719]}

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

Create PPO Agent with Recurrent Neural Networks

For this example load the predefined environment used for the Train DQN Agent to Balance Cart-Pole System example.

env = rlPredefinedEnv("CartPole-Discrete");

Get observation and action information. This environment has a continuous four-dimensional observation space (the positions and velocities of both cart and pole) and a discrete one-dimensional action space consisting on the application of two possible forces, -10N or 10N.

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

The actor and critic networks are initialized randomly. Ensure reproducibility by fixing the seed of the random generator.

rng(0)

PPO agents use a parametrized value function as a critic. To model the parametrized value function within the critic, use a recurrent neural network.

Define the network as an array of layer objects. To create a recurrent neural network, use a sequenceInputLayer as the input layer and include at least one lstmLayer.

criticNet = [
    sequenceInputLayer(prod(obsInfo.Dimension))
    fullyConnectedLayer(8)
    reluLayer
    lstmLayer(8)
    fullyConnectedLayer(1)
    ];

Convert to a dlnetwork object and display the number of learnable parameters.

criticNet = dlnetwork(criticNet);
summary(criticNet)

   Initialized: true

   Number of learnables: 593

   Inputs:
      1   'sequenceinput'   Sequence input with 4 dimensions

Create the critic using criticNet and the observation specification object. For more information, see rlValueFunction.

critic = rlValueFunction(criticNet,obsInfo);

Check the critic with a random observation input.

getValue(critic,{rand(obsInfo.Dimension)})

ans = single
    0.0017

Since the critic has a recurrent network, the actor must have a recurrent network too. Define the network as an array of layer objects.

actorNet = [
    sequenceInputLayer(prod(obsInfo.Dimension))
    fullyConnectedLayer(100)
    reluLayer
    lstmLayer(8)
    fullyConnectedLayer(numel(actInfo.Elements))
    softmaxLayer
    ];

Convert the network to a dlnetwork object and display the number of learnable parameters.

actorNet = dlnetwork(actorNet);
summary(actorNet)

   Initialized: true

   Number of learnables: 4k

   Inputs:
      1   'sequenceinput'   Sequence input with 4 dimensions

Create the actor using actorNet and the environment specification objects. For more information, see rlDiscreteCategoricalActor.

actor = rlDiscreteCategoricalActor(actorNet,obsInfo,actInfo);

Check the actor with a random observation input.

getAction(actor,{rand(obsInfo.Dimension)})

ans = 1x1 cell array
    {[-10]}

Set some training option for the critic.

criticOptions = rlOptimizerOptions( ...
    LearnRate=1e-2, ...
    GradientThreshold=1);

Set some training options for the actor.

actorOptions = rlOptimizerOptions( ...
    LearnRate=1e-3, ...
    GradientThreshold=1);

Create the agent options object.

agentOptions = rlPPOAgentOptions(...
    AdvantageEstimateMethod="finite-horizon", ...
    ClipFactor=0.1, ...
    CriticOptimizerOptions=criticOptions, ...
    ActorOptimizerOptions=actorOptions);

When recurrent neural networks are used, the MiniBatchSize property is the length of the learning trajectory.

agentOptions.MiniBatchSize

ans = 128

Create the agent using the actor and critic, as well as the agent options object.

agent = rlPPOAgent(actor,critic,agentOptions)

agent = 
  rlPPOAgent with properties:

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

Check your agent with a random observation input.

getAction(agent,rand(obsInfo.Dimension))

ans = 1x1 cell array
    {[-10]}

To evaluate the agent using sequential observations, use the sequence length (time) dimension. For example, obtain actions for a sequence of 9 observations.

[action,state] = getAction(agent, ...
    {rand([obsInfo.Dimension 1 9])});

Display the action corresponding to the seventh element of the observation.

action = action{1};
action(1,1,1,7)

ans = -10

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

Tips

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

• While tuning the learning rate of the actor network is necessary for PPO agents, it is not necessary for TRPO agents.

Version History

Introduced in R2019b