rlVectorQValueFunction

Vector Q-value function approximator with hybrid or discrete action space for reinforcement learning agents

Since R2022a

Description

This object implements a vector (also referred to as multi-output) Q-value function approximator that you can use as a critic with a hybrid or discrete action space for a reinforcement learning agent. A vector Q-value function (also known as vector action-value function) is a mapping between a pair consisting of an observation and a continuous action to a vector in which each element represents the expected discounted cumulative long-term reward when an agent executes the discrete action corresponding to the specific output number as well as the continuous action (if present) given as additional input, starting from the state corresponding to the observation (also given in input), and follows a given policy afterwards. A vector Q-value function critic for discrete action spaces therefore needs only the observation as input (as there is no continuous action). After you create an rlVectorQValueFunction critic, use it to create an agent such as rlQAgent, rlDQNAgent, rlSARSAAgent or an hybrid rlSACAgent. For more information on creating actors and critics, see Create Policies and Value Functions.

Creation

Syntax

critic = rlVectorQValueFunction(net,observationInfo,actionInfo)

critic = rlVectorQValueFunction({basisFcn,W0},observationInfo,actionInfo)

critic = rlVectorQValueFunction(___,Name=Value)

Description

critic = rlVectorQValueFunction(net,observationInfo,actionInfo) creates the multi-output Q-value function critic with a discrete or hybrid action space. Here, net is the deep neural network used as an approximation model. Its input layers must take in all the observation channels as well as and the continuous action channel (if present). The output layer must have as many elements as the number of possible discrete actions. When you use this syntax, the network input layers are automatically associated with the environment observation channels (and, if present, the continuous action channel) according to the dimensions specified in observationInfo (and if present, to the second action channel specified in actionInfo). This function sets the ObservationInfo and ActionInfo properties of critic to the observationInfo and actionInfo input arguments, respectively.

critic = rlVectorQValueFunction({basisFcn,W0},observationInfo,actionInfo) creates the multi-output Q-value function critic with a discrete or hybrid action space using a custom basis function as underlying approximation model. The first input argument is a two-element cell array whose first element is the handle basisFcn to a custom basis function and whose second element is the initial weight matrix W0. Here the basis function must take in all the observation channels as well as and the continuous action channel (if present), and W0 must have as many columns as the number of possible discrete actions. The function sets the ObservationInfo and ActionInfo properties of critic to the input arguments observationInfo and actionInfo, respectively.

critic = rlVectorQValueFunction(___,Name=Value) specifies names of the observation input layers (for network-based approximators) or sets the UseDevice property using one or more name-value arguments. Specifying the input layer names (and, for hybrid action spaces, the continuous action layer name) allows you explicitly associate the layers of your network approximator with specific environment channels. For all types of approximators, you can specify the device where computations for critic are executed, for example UseDevice="gpu".

Input Arguments

net — Deep neural network

array of Layer objects | layerGraph object | DAGNetwork object | SeriesNetwork object | dlNetwork object (preferred)

Deep neural network used as the underlying approximator within the critic, specified as one of the following:

• Array of Layer objects

• layerGraph object

• DAGNetwork object

• SeriesNetwork object

• dlnetwork object

Note

Among the different network representation options, dlnetwork is preferred, since it has built-in validation checks and supports automatic differentiation. If you pass another network object as an input argument, it is internally converted to a dlnetwork object. However, best practice is to convert other representations to dlnetwork explicitly before using it to create a critic or an actor for a reinforcement learning agent. You can do so using dlnet=dlnetwork(net), where net is any Deep Learning Toolbox™ neural network object. The resulting dlnet is the dlnetwork object that you use for your critic or actor. This practice allows a greater level of insight and control for cases in which the conversion is not straightforward and might require additional specifications.

The network must have as many input layers as the number of environment observation channels (plus one for hybrid action spaces). Specifically, there must be one input layer for each observation channel, and one additional input layer for the continuous action channel, if present. The output layer must have as many elements as the number of possible discrete actions. Each element of the output approximates the value of executing the corresponding discrete action starting from the current observation when also executing the continuous observation given as input (if present).

rlQValueFunction objects support recurrent deep neural networks.

The learnable parameters of the critic are the weights of the deep neural network. For a list of deep neural network layers, see List of Deep Learning Layers. For more information on creating deep neural networks for reinforcement learning, see Create Policies and Value Functions.

basisFcn — Custom basis function

function handle

Custom basis function, specified as a function handle to a user-defined MATLAB^® function. The user defined function can either be an anonymous function or a function on the MATLAB path. The output of the critic is the vector c = W'*B, where W is a matrix containing the learnable parameters, and B is the column vector returned by the custom basis function. Each element of c approximates the value of executing the corresponding discrete action starting from the current observation when also executing the continuous observation given as input (if present).

Your basis function must have the following signature.

B = myBasisFunction(obs1,obs2,...,obsN,actC)

Here, obs1 to obsN are inputs in the same order and with the same data type and dimensions as the channels defined in observationInfo. For hybrid action spaces, the additional input actC must have the same data type and dimension as the continuous action channel defined in the second element of actionInfo.

Example: @(obs1,obs2) [act(2)*obs1(1)^2; abs(obs2(5))]

W0 — Initial value of basis function weights

matrix

Initial value of the basis function weights W, specified as a matrix having as many rows as the length of the basis function output vector and as many columns as the number of possible discrete actions.

Name-Value Arguments

Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

Example: UseDevice="gpu"

ObservationInputNames — Network input layers names corresponding to the environment observation channels

string array | cell array of strings | cell array of character vectors

Network input layers names corresponding to the environment observation channels, specified as a string array or a cell array of strings or character vectors. The function assigns, in sequential order, each environment observation channel specified in observationInfo to each layer whose name is specified in the array assigned to this argument. Therefore, the specified network input layers, ordered as indicated in this argument, must have the same data type and dimensions as the observation channels, as ordered in observationInfo.

This name-value argument is supported only when the approximation model is a deep neural network.

Example: ObservationInputNames={"obsInLyr1_airspeed","obsInLyr2_altitude"}

ContinuousActionInputName — Network input layer name corresponding to the environment continuous action channel

string | character vector

Network input layer name corresponding to the environment continuous action channel, specified as a string array or a cell array of character vectors. When the action space has a continuous component, the function assigns the environment continuous action channel specified in the second element of actionInfo to the network input layer whose name is specified in the value assigned to this argument. Therefore, the specified network input layer must have the same data type and dimension as the continuous action channel defined in the second element of actionInfo.

This name-value argument is supported only when the approximation model is a deep neural network and when the action space is hybrid.

Example: ContinuousActionInputName="actCInLyr_Temperature"

Properties

ObservationInfo — Observation specifications

rlFiniteSetSpec object | rlNumericSpec object | array

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.

When you create the approximator object, the constructor function sets the ObservationInfo property to the input argument observationInfo.

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

Example: [rlNumericSpec([2 1]) rlFiniteSetSpec([3,5,7])]

ActionInfo — Action specification

vector containing one rlFiniteSetSpec followed by one rlNumericSpec object

Action specification, specified as a vector consisting of one rlFiniteSetSpec followed by one rlNumericSpec object. The action specification defines the properties of an environment action channel, such as its dimensions, data type, and name.

Note

For hybrid action spaces, you must have two action channels, the first one for the discrete part of the action, the second one for the continuous part of the action.

When you create the approximator object, the constructor function sets the ActionInfo property to the input argument actionInfo.

You can extract ActionInfo from an existing environment or agent using getActionInfo. You can also construct the specifications manually using rlFiniteSetSpec and rlNumericSpec.

Example: [rlFiniteSetSpec([-1 0 1]) rlNumericSpec([2 1])]

ActionInfo — Action specifications

rlFiniteSetSpec object

Action specifications, specified as an rlFiniteSetSpec object. This object defines the properties of the environment action channel, such as its dimensions, data type, and name.

Note

For this approximator object, only one action channel is allowed.

When you create the approximator object, the constructor function sets the ActionInfo property to the input argument actionInfo.

You can extract ActionInfo from an existing environment, approximator object, or agent using getActionInfo. You can also construct the specification manually using rlFiniteSetSpec.

Example: rlFiniteSetSpec([-1 0 1])

Normalization — Normalization method

"none" (default) | string array

Normalization method, returned as an array in which each element (one for each input channel defined in the observationInfo and actionInfo properties, in that order) is one of the following values:

• "none" — Do not normalize the input.

• "rescale-zero-one" — Normalize the input by rescaling it to the interval between 0 and 1. The normalized input Y is (U–Min)./(UpperLimit–LowerLimit), where U is the nonnormalized input. Note that nonnormalized input values lower than LowerLimit result in normalized values lower than 0. Similarly, nonnormalized input values higher than UpperLimit result in normalized values higher than 1. Here, UpperLimit and LowerLimit are the corresponding properties defined in the specification object of the input channel.

• "rescale-symmetric" — Normalize the input by rescaling it to the interval between –1 and 1. The normalized input Y is 2(U–LowerLimit)./(UpperLimit–LowerLimit) – 1, where U is the nonnormalized input. Note that nonnormalized input values lower than LowerLimit result in normalized values lower than –1. Similarly, nonnormalized input values higher than UpperLimit result in normalized values higher than 1. Here, UpperLimit and LowerLimit are the corresponding properties defined in the specification object of the input channel.

Note

When you specify the Normalization property of rlAgentInitializationOptions, normalization is applied only to the approximator input channels corresponding to rlNumericSpec specification objects in which both the UpperLimit and LowerLimit properties are defined. After you create the agent, you can use setNormalizer to assign normalizers that use any normalization method. For more information on normalizer objects, see rlNormalizer.

Example: "rescale-symmetric"

UseDevice — Computation device used for training and simulation

"cpu" (default) | "gpu"

Computation device used to perform operations such as gradient computation, parameter update and prediction during training and simulation, specified as either "cpu" or "gpu".

The "gpu" option requires both Parallel Computing Toolbox™ software and a CUDA^® enabled NVIDIA^® GPU. For more information on supported GPUs see GPU Computing Requirements (Parallel Computing Toolbox).

You can use gpuDevice (Parallel Computing Toolbox) to query or select a local GPU device to be used with MATLAB.

Note

Training or simulating an agent on a GPU involves device-specific numerical round-off errors. Because of these errors, you can get different results on a GPU and on a CPU for the same operation.

To speed up training by using parallel processing over multiple cores, you do not need to use this argument. Instead, when training your agent, use an rlTrainingOptions object in which the UseParallel option is set to true. For more information about training using multicore processors and GPUs for training, see Train Agents Using Parallel Computing and GPUs.

Example: "gpu"

Learnables — Learnable parameters of approximator object

cell array of dlarray objects

Learnable parameters of the approximator object, specified as a cell array of dlarray objects. This property contains the learnable parameters of the approximation model used by the approximator object.

Example: {dlarray(rand(256,4)),dlarray(rand(256,1))}

State — State of approximator object

cell array of dlarray objects

State of the approximator object, specified as a cell array of dlarray objects. For dlnetwork-based models, this property contains the Value column of the State property table of the dlnetwork model. The elements of the cell array are the state of the recurrent neural network used in the approximator (if any), as well as the state for the batch normalization layer (if used).

For model types that are not based on a dlnetwork object, this property is an empty cell array, since these model types do not support states.

Example: {dlarray(rand(256,1)),dlarray(rand(256,1))}

Object Functions

rlDQNAgent	Deep Q-network (DQN) reinforcement learning agent
rlQAgent	Q-learning reinforcement learning agent
rlSARSAAgent	SARSA reinforcement learning agent
getValue	Obtain estimated value from a critic given environment observations and actions
getMaxQValue	Obtain maximum estimated value over all possible actions from a Q-value function critic with discrete action space, given environment observations
evaluate	Evaluate function approximator object given observation (or observation-action) input data
getLearnableParameters	Obtain learnable parameter values from agent, function approximator, or policy object
setLearnableParameters	Set learnable parameter values of agent, function approximator, or policy object
setModel	Set approximation model in function approximator object
getModel	Get approximation model from function approximator object

Examples

Create Discrete Vector Q-Value Function Critic from Deep Neural Network

Create an observation specification object (or alternatively use getObservationInfo to extract the specification object from an environment). For this example, define the observation space as a continuous four-dimensional space, so that there is a single observation channel that carries a column vector containing four doubles.

obsInfo = rlNumericSpec([4 1]);

Create a discrete (finite set) action specification object (or alternatively use getActionInfo to extract the specification object from an environment with a discrete action space). For this example, define the action space as a finite set consisting of three possible actions (labeled 7, 5, and 3).

actInfo = rlFiniteSetSpec([7 5 3]);

A discrete vector Q-value function takes only the observation as input and returns as output a single vector with as many elements as the number of possible discrete actions. The value of each output element represents the expected discounted cumulative long-term reward for taking the action corresponding to the element number, from the state corresponding to the current observation, and following the policy afterwards.

To model the parametrized vector Q-value function within the critic, use a neural network with one input layer (receiving the content of the observation channel, as specified by obsInfo) and one output layer (returning the vector of values for all the possible actions, as specified by actInfo).

Define the network as an array of layer objects, and get the dimension of the observation space and the number of possible actions from the environment specification objects.

net = [  
    featureInputLayer(obsInfo.Dimension(1))
    fullyConnectedLayer(16)
    reluLayer
    fullyConnectedLayer(16)
    reluLayer
    fullyConnectedLayer(numel(actInfo.Elements)) 
    ];

Convert the network to a dlnetwork object, and display the number of weights.

net = dlnetwork(net);
summary(net)

   Initialized: true

   Number of learnables: 403

   Inputs:
      1   'input'   4 features

Create the critic using the network, as well as the observation and action specification objects.

critic = rlVectorQValueFunction(net,obsInfo,actInfo)

critic = 
  rlVectorQValueFunction with properties:

    ObservationInfo: [1x1 rl.util.rlNumericSpec]
         ActionInfo: [1x1 rl.util.rlFiniteSetSpec]
      Normalization: "none"
          UseDevice: "cpu"
         Learnables: {6x1 cell}
              State: {0x1 cell}

To check your critic, use getValue to return the values of the discrete actions depending on a random observation, using the current network weights. There is one value for each of the three possible actions.

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

v = 3x1 single column vector

    0.0761
   -0.5906
    0.2072

You can now use the critic to create an agent for the environment described by the given specification objects. Examples of agents that can use a discrete vector Q-value function critic, are rlQAgent, rlDQNAgent, rlSARSAAgent and discrete rlSACAgent.

For more information on creating approximator objects such as actors and critics, see Create Policies and Value Functions.

Create Hybrid Vector Q-Value Function Critic from Deep Neural Network

Create an observation specification object (or alternatively use getObservationInfo to extract the specification object from an environment). For this example, define the observation space as a continuous four-dimensional space, so that there is a single observation channel that carries a column vector containing four doubles.

obsInfo = rlNumericSpec([4 1]);

Create a hybrid action specification object (or alternatively use getActionInfo to extract the specification object from an environment with a hybrid action space). For this example, define the action space as having the discrete channel carrying a scalar that can take three values (labeled -1, 0, and 1), and the continuous channel carrying a vector in which each of the three elements can vary between -10 and 10. For an hybrid action space, the discrete channel must always be the first one, and the continuous channel must be the second one.

actInfo = [ 
    rlFiniteSetSpec([-1 0 1])
    rlNumericSpec([3 1], ...
         UpperLimit= 10*ones(3,1), ...
         LowerLimit=-10*ones(3,1) )
    ];

An hybrid vector Q-value function takes the observations and the continuous action as inputs and returns as output a single vector with as many elements as the number of possible discrete actions. The value of each output element represents the expected discounted cumulative long-term reward when an agent executes the discrete action corresponding to the specific output number as well as the continuous action given in input, starting from the state corresponding to the observation (also given in input), and follows the given policy afterwards.

To model the parametrized vector Q-value function within the critic, use a neural network with two input layers (receiving the content of the observation and continuous action channels, as specified by obsInfo and the second element of actInfo), and one output layer (returning the vector of values for the possible discrete actions, as specified by the first element of actInfo).

Define each network path as an array of layer objects, and get the dimension of the observation space and the number of possible actions from the environment specification objects.

Create the observation input path.

obsPath = [ 
    featureInputLayer( ...
        prod(obsInfo.Dimension), ...
        Name="obsInLyr")
    fullyConnectedLayer(prod(obsInfo.Dimension))
    reluLayer(Name="obsPthOutLyr")
    ];

Create the continuous action input path.

actCPath = [ 
    featureInputLayer(prod(actInfo(2).Dimension), ...
        Name="actCInLyr")
    fullyConnectedLayer( ...
        prod(actInfo(2).Dimension))
    reluLayer(Name="actCPthOutLyr")
    ];

Create common output path. Concatenate inputs along the first available dimension.

valuesPath = [ 
    concatenationLayer(1,2,Name="valPthInLyr")
    fullyConnectedLayer(numel(actInfo(1).Elements))
    reluLayer
    fullyConnectedLayer( ...
        numel(actInfo(1).Elements), ...
        Name="valPthOutLyr")
    ];

Assemble dlnetwork object.

net = dlnetwork;
net = addLayers(net,obsPath);
net = addLayers(net,actCPath);
net = addLayers(net,valuesPath);

Connect layers.

net = connectLayers(net,"obsPthOutLyr","valPthInLyr/in1");
net = connectLayers(net,"actCPthOutLyr","valPthInLyr/in2");

Plot network.

plot(net)

Initialize network.

net = initialize(net);

Display the number of learnable parameters.

summary(net)

   Initialized: true

   Number of learnables: 68

   Inputs:
      1   'obsInLyr'    4 features
      2   'actCInLyr'   3 features

Create the critic using the network, as well as the observation and action specification objects. When you use this syntax, the network input layers are automatically associated with the environment observation channels according to the dimensions specified in obsInfo.

critic = rlVectorQValueFunction(net,obsInfo,actInfo)

critic = 
  rlVectorQValueFunction with properties:

    ObservationInfo: [1x1 rl.util.rlNumericSpec]
         ActionInfo: [2x1 rl.util.RLDataSpec]
      Normalization: ["none"    "none"]
          UseDevice: "cpu"
         Learnables: {8x1 cell}
              State: {0x1 cell}

To check your critic, use getValue to return the values the discrete actions depending on a random observation and a random continuous action, using the current network weights. There is one value for each of the three possible actions.

v = getValue(critic, ...
    {rand(obsInfo.Dimension),rand(actInfo(2).Dimension)})

v = 3x1 single column vector

    0.5603
   -0.4927
    0.3953

You can now use the critic to create an hybrid rlSACAgent.

For more information on creating approximator objects such as actors and critics, see Create Policies and Value Functions.

Create Hybrid Vector Q-Value Function Critic from Deep Neural Network Specifying Layer Names

Create an observation specification object (or alternatively use getObservationInfo to extract the specification object from an environment). For this example, define the observation space as an hybrid (that is mixed discrete-continuous) space with the discrete channel carrying a scalar that can be either 0 or 1, and the second one being a vector over a continuous two-dimensional space.

obsInfo = [
    rlFiniteSetSpec([0 1])
    rlNumericSpec([2 1]) 
    ];

Create a specification object for an hybrid action space (or alternatively use getActionInfo to extract the specification object from an environment with a hybrid action space). For this example, define the action space as an hybrid space with the discrete channel carrying a scalar that can only be -1, 0, or 1, and with the continuous channel carrying a two-element vector. Note that when defining hybrid action spaces, you must define the discrete action channel as the first one and the continuous action channel as the second one.

actInfo = [ 
    rlFiniteSetSpec([-1 0 1])
    rlNumericSpec([2 1])
    ];

A hybrid vector Q-value function takes the observations and the continuous action as inputs and returns as output a single vector with as many elements as the number of possible discrete actions. The value of each output element represents the expected discounted cumulative long-term reward for taking the discrete action corresponding to the element number and the continuous action given as input, from the state corresponding to the current observations, and following the policy afterwards.

To model the parametrized vector Q-value function within the critic, use a neural network with three input layers and one output layer.

• The input layers receive the content of the two observation channels, as specified by obsInfo as well as the content of the continuous action channel, as specified by the second element of actInfo. Therefore each input layer must have the same dimensions as the corresponding observation channel and an extra input layer must have the same dimensions of the continuous action channel, as specified by the second element of actInfo.

• Each element of the output layer returns the value of the corresponding action. Therefore the output layer must have as many elements as the number of possible discrete actions, as specified by the first element of actInfo.

Define each network path as an array of layer objects, and get the dimension of the observation space and the number of possible actions from the environment specification objects. Name the network inputs obs1InLyr , obs2InLyr and actCInLyr (so you can later explicitly associate it with the observation input channel).

Create the first observation input path.

obs1Path = [ 
    featureInputLayer( ...
        prod(obsInfo(1).Dimension), ...
        Name="obs1InLyr")            
    fullyConnectedLayer( ...
        prod(obsInfo(1).Dimension))
    reluLayer(Name="obs1PthOutLyr")
    ];

Create the second observation input path.

obs2Path = [ 
    featureInputLayer( ...
        prod(obsInfo(2).Dimension), ...
        Name="obs2InLyr")
    fullyConnectedLayer( ...
        prod(obsInfo(2).Dimension))
    reluLayer(Name="obs2PthOutLyr")
    ];

Create the continuous action input path.

actCPath = [ 
    featureInputLayer( ...
        prod(actInfo(2).Dimension), ...
        Name="actCInLyr")
    fullyConnectedLayer( ...
        prod(actInfo(2).Dimension))
    reluLayer(Name="actCPthOutLyr")
    ];

Create common output path. Concatenate inputs along the first available dimension.

valuePath = [ 
    concatenationLayer(1,3,Name="valPthInLyr")
    fullyConnectedLayer(numel(actInfo(1).Elements))
    reluLayer
    fullyConnectedLayer( ...
        numel(actInfo(1).Elements), ...
        Name="valPthOutLyr")
    ];

Assemble dlnetwork object.

net = dlnetwork;
net = addLayers(net,obs1Path);
net = addLayers(net,obs2Path);
net = addLayers(net,actCPath);
net = addLayers(net,valuePath);

Connect layers.

net = connectLayers(net,"obs1PthOutLyr","valPthInLyr/in1");
net = connectLayers(net,"obs2PthOutLyr","valPthInLyr/in2");
net = connectLayers(net,"actCPthOutLyr","valPthInLyr/in3");

Plot network.

plot(net)

Initialize network.

net = initialize(net);

Display the number of learnable parameters.

summary(net)

   Initialized: true

   Number of learnables: 44

   Inputs:
      1   'obs1InLyr'   1 features
      2   'obs2InLyr'   2 features
      3   'actCInLyr'   2 features

Create the critic using the network, the observations specification object, and the names of the network input layers. In this case, attempting to create a critic without using layer names will result in an error, because the dimension of the action channel is the same as the one of the continuous action channel, so the software is unable to unequivocally associate network layers with environment channels.

The specified network input layer, obsInLyr, is associated with the environment observation, and therefore must have the same data type and dimension as the observation channel specified in obsInfo.

critic = rlVectorQValueFunction(net,obsInfo,actInfo, ...
    ObservationInputNames={"obs1InLyr","obs2InLyr"}, ...
    ContinuousActionInputName="actCInLyr")

critic = 
  rlVectorQValueFunction with properties:

    ObservationInfo: [2x1 rl.util.RLDataSpec]
         ActionInfo: [2x1 rl.util.RLDataSpec]
      Normalization: ["none"    "none"    "none"]
          UseDevice: "cpu"
         Learnables: {10x1 cell}
              State: {0x1 cell}

To check your critic, use getValue to return the values the discrete actions depending on a random observation and a random continuous action, using the current network weights. There is one value for each of the three possible actions.

v = getValue(critic,{ ...
    rand(obsInfo(1).Dimension), ...
    rand(obsInfo(2).Dimension), ...
    rand(actInfo(2).Dimension) })

v = 3x1 single column vector

    0.5075
   -0.2680
    0.0327

Calculate the values of the discrete actions for a batch of five independent observations and continuous actions.

v = getValue(critic,{ ...
    rand([obsInfo(1).Dimension 5]), ...
    rand([obsInfo(2).Dimension 5]), ...
    rand([actInfo(2).Dimension 5]) })

v = 3x5 single matrix

    1.1344    0.4129    0.9857    0.7525    1.1961
   -0.4883   -0.8504   -0.8007   -0.9811   -0.8754
    0.1460   -0.7302   -0.1208   -0.4614   -0.0832

You can now use the critic to create an hybrid rlSACAgent.

For more information on creating approximator objects such as actors and critics, see Create Policies and Value Functions.

Create Hybrid Q-Value Function Critic from Custom Basis Function

Create an observation specification object (or alternatively use getObservationInfo to extract the specification object from an environment). For this example, define the observation space as consisting of two channels, the first carrying a two-by-two continuous matrix and the second carrying scalar that can assume only two values, 0 and 1.

obsInfo = [
    rlNumericSpec([2 2]) 
    rlFiniteSetSpec([0 1])
    ];

Create a hybrid action specification object (or alternatively use getActionInfo to extract the specification object from an environment with a hybrid action space). For this example, define a hybrid action space in which the discrete part is a vector that can have three possible values: [1 2], [3 4], and [5 6], while the continuous part is a two-dimensional continuous vector in which each element can very between -5 and 5.

actInfo = [ 
    rlFiniteSetSpec({[1 2],[3 4],[5 6]})
    rlNumericSpec([2 1], ...
         UpperLimit= 5*ones(2,1), ...
         LowerLimit=-5*ones(2,1) )
   ];      

An hybrid vector Q-value function takes observations and the continuous action as inputs and returns as output a single vector with as many elements as the number of possible discrete actions. The value of each element of the output vector represents the expected discounted cumulative long-term reward when an agent executes the discrete action corresponding to the specific output number as well as the continuous action given in input, starting from the state corresponding to the observation (also given in input), and follows the given policy afterwards.

To model the parametrized vector Q-value function within the critic, use a custom basis function with three inputs and one output.

• The first two inputs receive the content of the environment observation channels, as specified by obsInfo (for this example, a matrix and a scalar, respectively).

• The third input receives the content of the continuous action channel, as specified by the second element of actInfo, (for this example, a vector).

• The output is a vector with four elements (you can have as many elements as needed by your application).

Create a function that returns a vector of four elements, given the observations and the continuous action as inputs.

Here, the third dimension is the batch dimension. For each element of the batch dimension, the output of the basis function is a vector with four elements. Each output element can be any combination of the three inputs, depending on your application.

myBasisFcn = @(obsA,obsB,actC) [
    obsA(1,1,:)+obsB(1,1,:).^2-actC(1,1,:);
    obsA(2,1,:)-obsB(1,1,:).^2-actC(2,1,:);
    obsA(1,2,:).^2+obsB(1,1,:)+actC(1,1,:);
    obsA(2,2,:).^2-obsB(1,1,:)+actC(2,1,:);
    ];

For each element of the batch, the output of the critic is the vector c = W'*myBasisFcn(obsA,obsB,actC), where W is a weight matrix which must have as many rows as the length of the basis function output and as many columns as the number of possible discrete actions, as specified by the first element of actInfo.

Each element of c represents the expected cumulative long term reward when an agent executes the discrete action corresponding to number of the element in c, as well as the continuous action given in input, starting from the state corresponding to the observations (also given in input), and follows the given policy afterwards.

The elements of W are the learnable parameters.

Define an initial parameter matrix.

W0 = rand(4,3);

Create the critic. The first argument is a two-element cell containing both the handle to the custom function and the initial parameter matrix. The second and third arguments are, respectively, the observation and action specification objects.

critic = rlVectorQValueFunction({myBasisFcn,W0},obsInfo,actInfo)

critic = 
  rlVectorQValueFunction with properties:

    ObservationInfo: [2x1 rl.util.RLDataSpec]
         ActionInfo: [2x1 rl.util.RLDataSpec]
      Normalization: ["none"    "none"    "none"]
          UseDevice: "cpu"
         Learnables: {[3x4 dlarray]}
              State: {}

To check your critic, use getValue to return the values of a random observation, using the current parameter matrix. The function returns one value for each of the three possible actions.

v = getValue(critic,{rand(2,2),0,rand(2,1)})

v = 3x1 single column vector

    1.3972
    1.0399
    1.5082

Note that the critic does not enforce the set constraint for the discrete set elements.

v = getValue(critic,{rand(2,2),-1,rand(2,1)})

v = 3x1 single column vector

    2.2854
    1.8893
    2.5565

Obtain values for a random batch of 10 observations.

v = getValue(critic,{ ...
    rand([obsInfo(1).Dimension 10]), ...
    rand([obsInfo(2).Dimension 10]) ...
    rand([actInfo(2).Dimension 10]) ...
    });

Display the values corresponding to the seventh element of the observation batch.

v(:,7)

ans = 3x1 single column vector

    0.8710
    0.7387
    0.9314

You can now use the critic to create an agent for the environment described by the given specification objects.

You can now use the critic to create an hybrid rlSACAgent.

For more information on creating approximator objects such as actors and critics, see Create Policies and Value Functions.

Create Discrete Vector Q-Value Function Critic from Recurrent Neural Network

Create an environment and obtain observation and action specification objects.

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

A discrete vector Q-value function takes only the observation as input and returns as output a single vector with as many elements as the number of possible actions. The value of each output element represents the expected discounted cumulative long-term reward for taking the action from the state corresponding to the current observation, and following the policy afterwards.

To model the parametrized vector Q-value function within the critic, use a recurrent neural network with one input layer (receiving the content of the observation channel, as specified by obsInfo) and one output layer (returning the vector of values for all the possible actions, as specified by actInfo).

Define the network as an array of layer objects, and get the dimension of the observation space and the number of possible actions from the environment specification objects. To create a recurrent network, use a sequenceInputLayer as the input layer (with size equal to the number of dimensions of the observation channel) and include at least one lstmLayer.

net = [
    sequenceInputLayer(obsInfo.Dimension(1))
    fullyConnectedLayer(50)
    reluLayer
    lstmLayer(20)
    fullyConnectedLayer(20)
    reluLayer
    fullyConnectedLayer(numel(actInfo.Elements)) 
    ];

Convert the network to a dlnetwork object, and display the number of weights.

net = dlnetwork(net);
summary(net)

   Initialized: true

   Number of learnables: 6.3k

   Inputs:
      1   'sequenceinput'   Sequence input with 4 dimensions

Create the critic using the network, as well as the observation and action specification objects.

critic = rlVectorQValueFunction(net, ...
    obsInfo,actInfo);

To check your critic, use getValue to return the value of a random observation and action, using the current network weights.

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

v = 2x1 single column vector

    0.0136
    0.0067

You can use dot notation to extract and set the current state of the recurrent neural network in the critic.

critic.State

ans=2×1 cell array
    {20x1 dlarray}
    {20x1 dlarray}

critic.State = { 
    -0.1*dlarray(rand(20,1))
     0.1*dlarray(rand(20,1))
     };

To evaluate the critic using sequential observations, use the sequence length (time) dimension. For example, obtain actions for 5 independent sequences each one consisting of 9 sequential observations.

[value,state] = getValue(critic, ...
    {rand([obsInfo.Dimension 5 9])});

Display the value corresponding to the seventh element of the observation sequence in the fourth sequence.

value(1,4,7)

ans = single

0.0382

Display the updated state of the recurrent neural network.

state

state=2×1 cell array
    {20x5 single}
    {20x5 single}

You can now use the critic to create an agent for the environment described by the given specification objects. Examples of agents that can use a vector Q-value function critic, are rlQAgent, rlDQNAgent, rlSARSAAgent, and discrete action space rlSACAgent.

For more information on creating approximator objects such as actors and critics, see Create Policies and Value Functions.

Version History

Introduced in R2022a