Reinforcement Learning Action Set: Syntax

reinforcementLearn.rlTrainPolicyGradient <result=results> <status=rc> /

={{

act="IDENTITY" | "RELU" | "SIGMOID" | "TANH",

n=integer,

size=integer,

stride=integer,

type="CONV" | "FC"

}, {...}},

actorOptimizer={method="ADAM" | "SGD", method-specific-parameters},

caslib="string",

compress=TRUE | FALSE,

indexVars={"variable-name-1" <, "variable-name-2", ...>},

label="string",

lifetime=64-bit-integer,

maxMemSize=64-bit-integer,

memoryFormat="DVR" | "INHERIT" | "STANDARD",

name="table-name",

onDemand=TRUE | FALSE,

promote=TRUE | FALSE,

replace=TRUE | FALSE,

replication=integer,

threadBlockSize=64-bit-integer,

timeStamp="string",

where={"string-1" <, "string-2", ...>}

checkPointFreq=integer,

caslib="string",

computedOnDemand=TRUE | FALSE,

computedVars

={{

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

}, {...}},

computedVarsProgram="string",

dataSourceOptions={key-1=any-list-or-data-type-1 <, key-2=any-list-or-data-type-2, ...>},

groupBy

={{

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

}, {...}},

groupByMode="NOSORT" | "REDISTRIBUTE",

importOptions={fileType="ANY" | "AUDIO" | "AUTO" | "BASESAS" | "CSV" | "DELIMITED" | "DOCUMENT" | "DTA" | "ESP" | "EXCEL" | "FMT" | "HDAT" | "IMAGE" | "JMP" | "LASR" | "PARQUET" | "SOUND" | "SPSS" | "VIDEO" | "XLS", fileType-specific-parameters},

name="table-name",

onDemand=TRUE | FALSE,

orderBy

={{

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

}, {...}},

singlePass=TRUE | FALSE,

vars

={{

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

}, {...}},

where="where-expression",

whereTable

casLib="string"

dataSourceOptions={adls_noreq-parameters | bigquery-parameters | cas_noreq-parameters | clouddex-parameters | db2-parameters | dnfs-parameters | esp-parameters | fedsvr-parameters | gcs_noreq-parameters | greenplum-parameters | hadoop-parameters | hana-parameters | hdfs-parameters | impala-parameters | jdbc-parameters | mongodb-parameters | mysql-parameters | netezza-parameters | odbc-parameters | oracle-parameters | path-parameters | postgres-parameters | redshift-parameters | s3-parameters | sapiq-parameters | sforce-parameters | singlestore_standard-parameters | snowflake-parameters | spark-parameters | spde-parameters | sqlserver-parameters | ss_noreq-parameters | teradata-parameters | vertica-parameters | yellowbrick-parameters}

name="table-name"

vars

={{

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

}, {...}}

where="where-expression"

}

caslib="string",

compress=TRUE | FALSE,

indexVars={"variable-name-1" <, "variable-name-2", ...>},

label="string",

lifetime=64-bit-integer,

maxMemSize=64-bit-integer,

memoryFormat="DVR" | "INHERIT" | "STANDARD",

name="table-name",

onDemand=TRUE | FALSE,

promote=TRUE | FALSE,

replace=TRUE | FALSE,

replication=integer,

threadBlockSize=64-bit-integer,

timeStamp="string",

where={"string-1" <, "string-2", ...>}

={{

act="IDENTITY" | "RELU" | "SIGMOID" | "TANH",

n=integer,

size=integer,

stride=integer,

type="CONV" | "FC"

}, {...}},

criticOptimizer={method="ADAM" | "SGD", method-specific-parameters},

environment={type="BUILTIN" | "REMOTE", type-specific-parameters},

gamma=double,

deterministic=TRUE | FALSE,

device=integer,

enable=TRUE | FALSE

maxNormOfGradient=double,

caslib="string",

compress=TRUE | FALSE,

indexVars={"variable-name-1" <, "variable-name-2", ...>},

label="string",

lifetime=64-bit-integer,

maxMemSize=64-bit-integer,

memoryFormat="DVR" | "INHERIT" | "STANDARD",

name="table-name",

onDemand=TRUE | FALSE,

promote=TRUE | FALSE,

replace=TRUE | FALSE,

replication=integer,

threadBlockSize=64-bit-integer,

timeStamp="string",

where={"string-1" <, "string-2", ...>}

nThreads=integer,

numEpisodes=integer,

numTestEpisodes=integer,

pgAlgorithm={type="ACTORCRITIC" | "REINFORCE" | "REINFORCERTG", type-specific-parameters},

seed=double,

testInterval=integer

;

indicates a required parameter

Summary: Input and Output Tables

If a row includes a subparameter, you can specify the name, caslib, and so on in the subparameter. Otherwise, you can specify the name, caslib, and so on in the parameter.

Parameters for Reading Input Tables
Parameter	Subparameter	Description
checkPointIn	—	Specifies a table that includes a saved model for warm starting.

Parameters for Creating Output Tables
Parameter	Subparameter	Description
checkPointBestOut	—	Specifies a table that to save the best observed model.
checkPointOut	—	Specifies a table that is used to save the model.
required parametermodelOut	—	Specifies a table that is used to save model structure and weights.

Parameter Descriptions

* actorModel={{nnParams-1} <, {nnParams-2}, ...>}

Specifies the properties of the layers of the neural net used to build the actor network.

The nnParams value can be one or more of the following:

act="IDENTITY" | "RELU" | "SIGMOID" | "TANH"

Activation function.

Default	RELU

n=integer

Number of neurons.

Range	1–100000

size=integer

Size (width and height) of kernel (for convolution layer) or window (for pooling layer).

Range	1–1000

stride=integer

Stride of kernel (for convolution layer) or window (for pooling layer).

Default	1
Range	1–10000

type="CONV" | "FC"

Layer type.

Default	FC

actorOptimizer={method="ADAM" | "SGD", method-specific-parameters}

The optimization method for the actor model: SGD or ADAM.

Alias	actorOptimization

The value that you specify for the method parameter determines the other parameters that apply.

checkPointBestOut={casouttable}

Specifies a table that to save the best observed model.

For more information about specifying the checkPointBestOut parameter, see the common casouttable parameter.

checkPointFreq=integer

Specifies how often to save training parameters.

Range	1–100000000

checkPointIn={castable}

Specifies a table that includes a saved model for warm starting.

For more information about specifying the checkPointIn parameter, see the common castable parameter.

checkPointOut={casouttable}

Specifies a table that is used to save the model.

For more information about specifying the checkPointOut parameter, see the common casouttable parameter.

criticModel={{nnParams-1} <, {nnParams-2}, ...>}

Specifies the properties of the layers of the neural net used to build the critic network. If this value is left unspecified, the value from actorModel is copied.

The nnParams value can be one or more of the following:

act="IDENTITY" | "RELU" | "SIGMOID" | "TANH"

Activation function.

Default	RELU

n=integer

Number of neurons.

Range	1–100000

size=integer

Size (width and height) of kernel (for convolution layer) or window (for pooling layer).

Range	1–1000

stride=integer

Stride of kernel (for convolution layer) or window (for pooling layer).

Default	1
Range	1–10000

type="CONV" | "FC"

Layer type.

Default	FC

criticOptimizer={method="ADAM" | "SGD", method-specific-parameters}

The optimization method for the critic model: SGD or ADAM. If this value is left unspecified, the value from actorOptimizer is copied.

Alias	criticOptimization

The value that you specify for the method parameter determines the other parameters that apply.

* environment={type="BUILTIN" | "REMOTE", type-specific-parameters}

The reinforcement learning environment.

Alias	env

The value that you specify for the type parameter determines the other parameters that apply.

gamma=double

Discount factor used to compute long term value.

Default	0.99
Range	0–1

gpu={gpuOptions}

Specifies GPU related options.

The gpuOptions value can be one or more of the following:

deterministic=TRUE | FALSE

Specifies that the action will use deterministic GPU functions (may be slower).

Default	FALSE

device=integer

Specifies a specific GPU device id.

Minimum value	0

enable=TRUE | FALSE

Specifies that action will use a GPU to perform model calculations.

Default	FALSE

maxNormOfGradient=double

Adjust the magnitude of the gradient so that it is less than this value.

Range	1–100000000

* modelOut={casouttable}

Specifies a table that is used to save model structure and weights.

For more information about specifying the modelOut parameter, see the common casouttable parameter.

nThreads=integer

Number of CPU threads to use.

Range	0–1024

numEpisodes=integer

Number of episodes to train before terminating.

Default	10
Range	1–100000000

numTestEpisodes=integer

Specifies number of episodes to run when running test.

Default	100
Range	1–100000000

* pgAlgorithm={type="ACTORCRITIC" | "REINFORCE" | "REINFORCERTG", type-specific-parameters}

The policy gradient algorithm type, REINFORCE, REINFORCERTG, ActorCritic, A2C

Alias	pgAlg

The value that you specify for the type parameter determines the other parameters that apply.

seed=double

Seed for random number generator.

Minimum value	0

testInterval=integer

Specifies how often to run test.

Default	100
Range	1–100000000

Parameters for method="ADAM"

beta1=double

The exponential decay rate for the first moment estimates.

Default	0.9
Range	0–1

beta2=double

The exponential decay rate for the second moment estimates.

Default	0.999
Range	0–1

epsilon=double

The value to add into the second moment of the gradient in the denominator to improve numerical stability.

Alias	eps
Default	1E-08
Range	0–1

learningRate=double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

miniBatchSize=integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

regL2=double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

useAMSGrad=TRUE | FALSE

Use AMSGrad variant of ADAM. See paper: On the Convergence of Adam and Beyond.

Default	FALSE

Parameters for method="SGD"

damping=double

Damping for momentum.

Default	0
Range	0–1

learningRate=double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

miniBatchSize=integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

momentum=double

Specifies the value for momentum for SGD.

Default	0
Range	0–1

regL2=double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

useNesterov=TRUE | FALSE

Enables Nesterov momentum.

Default	FALSE

Parameters for method="ADAM"

beta1=double

The exponential decay rate for the first moment estimates.

Default	0.9
Range	0–1

beta2=double

The exponential decay rate for the second moment estimates.

Default	0.999
Range	0–1

epsilon=double

The value to add into the second moment of the gradient in the denominator to improve numerical stability.

Alias	eps
Default	1E-08
Range	0–1

learningRate=double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

miniBatchSize=integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

regL2=double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

useAMSGrad=TRUE | FALSE

Use AMSGrad variant of ADAM. See paper: On the Convergence of Adam and Beyond.

Default	FALSE

Parameters for method="SGD"

damping=double

Damping for momentum.

Default	0
Range	0–1

learningRate=double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

miniBatchSize=integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

momentum=double

Specifies the value for momentum for SGD.

Default	0
Range	0–1

regL2=double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

useNesterov=TRUE | FALSE

Enables Nesterov momentum.

Default	FALSE

Parameters for type="BUILTIN"

* name="string"

Parameters for type="REMOTE"

imageStack=integer

Specifies number of consecutive observations to be stacked as an input observation.

Default	1
Range	1–10000

* name="string"

The environment name used in RL.

render=TRUE | FALSE

Whether render the environment display or not.

Default	FALSE

renderFreq=integer

Specifies how frequently the environment is rendered.

Alias	renderFrequency
Default	1
Range	1–10000000

renderSleep=double

Time to sleep (in seconds) between rendering frames.

Default	0

seed=integer

Seed of the environment.

Default	0

* url="string"

The url:port used for in RL training.

Default	"127.0.0.1:8080"

wrapper={"string-1" <, "string-2", ...>}

A list of strings that specifies the wrappers that will wrap the environment. The wrappers are used to preprocess the environment observations and rewards.

Parameters for type="ACTORCRITIC"

entropyCoef=double

Specifies the value for the entropy coefficient in policy gradient algorithms.

Default	0.01
Range	0–1

GAELambda=double

Specifies the value of the lambda in generalized advantage estimation (GAE).

Default	0.9
Range	0–1

nRollouts=integer

Specifies the number of rollouts to generate data.

Default	1
Range	1–1024

useGAELoss=TRUE | FALSE

Use GAE Loss.

Default	FALSE

Parameters for type="REINFORCE"

entropyCoef=double

Specifies the value for the entropy coefficient in policy gradient algorithms.

Default	0.01
Range	0–1

nRollouts=integer

Specifies the number of rollouts to generate data.

Default	1
Range	1–1024

Parameters for type="REINFORCERTG"

entropyCoef=double

Specifies the value for the entropy coefficient in policy gradient algorithms.

Default	0.01
Range	0–1

nRollouts=integer

Specifies the number of rollouts to generate data.

Default	1
Range	1–1024

rlTrainPolicyGradient Action

Do reinforcement learning training using Policy Gradient algorithm.

Alias	rlTrainPG

Lua Syntax

results, info = s:reinforcementLearn_rlTrainPolicyGradient{

={{

act="IDENTITY" | "RELU" | "SIGMOID" | "TANH",

n=integer,

size=integer,

stride=integer,

type="CONV" | "FC"

}, {...}},

actorOptimizer={method="ADAM" | "SGD", method-specific-parameters},

caslib="string",

compress=true | false,

indexVars={"variable-name-1" <, "variable-name-2", ...>},

label="string",

lifetime=64-bit-integer,

maxMemSize=64-bit-integer,

memoryFormat="DVR" | "INHERIT" | "STANDARD",

name="table-name",

onDemand=true | false,

promote=true | false,

replace=true | false,

replication=integer,

threadBlockSize=64-bit-integer,

timeStamp="string",

where={"string-1" <, "string-2", ...>}

checkPointFreq=integer,

caslib="string",

computedOnDemand=true | false,

computedVars

={{

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

}, {...}},

computedVarsProgram="string",

dataSourceOptions={key-1=any-list-or-data-type-1 <, key-2=any-list-or-data-type-2, ...>},

groupBy

={{

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

}, {...}},

groupByMode="NOSORT" | "REDISTRIBUTE",

name="table-name",

onDemand=true | false,

orderBy

={{

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

}, {...}},

singlePass=true | false,

vars

={{

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

}, {...}},

where="where-expression",

whereTable

casLib="string"

name="table-name"

vars

={{

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

}, {...}}

where="where-expression"

}

caslib="string",

compress=true | false,

indexVars={"variable-name-1" <, "variable-name-2", ...>},

label="string",

lifetime=64-bit-integer,

maxMemSize=64-bit-integer,

memoryFormat="DVR" | "INHERIT" | "STANDARD",

name="table-name",

onDemand=true | false,

promote=true | false,

replace=true | false,

replication=integer,

threadBlockSize=64-bit-integer,

timeStamp="string",

where={"string-1" <, "string-2", ...>}

={{

act="IDENTITY" | "RELU" | "SIGMOID" | "TANH",

n=integer,

size=integer,

stride=integer,

type="CONV" | "FC"

}, {...}},

criticOptimizer={method="ADAM" | "SGD", method-specific-parameters},

environment={type="BUILTIN" | "REMOTE", type-specific-parameters},

gamma=double,

deterministic=true | false,

device=integer,

enable=true | false

maxNormOfGradient=double,

caslib="string",

compress=true | false,

indexVars={"variable-name-1" <, "variable-name-2", ...>},

label="string",

lifetime=64-bit-integer,

maxMemSize=64-bit-integer,

memoryFormat="DVR" | "INHERIT" | "STANDARD",

name="table-name",

onDemand=true | false,

promote=true | false,

replace=true | false,

replication=integer,

threadBlockSize=64-bit-integer,

timeStamp="string",

where={"string-1" <, "string-2", ...>}

nThreads=integer,

numEpisodes=integer,

numTestEpisodes=integer,

pgAlgorithm={type="ACTORCRITIC" | "REINFORCE" | "REINFORCERTG", type-specific-parameters},

seed=double,

testInterval=integer

}

indicates a required parameter

Summary: Input and Output Tables

If a row includes a subparameter, you can specify the name, caslib, and so on in the subparameter. Otherwise, you can specify the name, caslib, and so on in the parameter.

Parameters for Reading Input Tables
Parameter	Subparameter	Description
checkPointIn	—	Specifies a table that includes a saved model for warm starting.

Parameters for Creating Output Tables
Parameter	Subparameter	Description
checkPointBestOut	—	Specifies a table that to save the best observed model.
checkPointOut	—	Specifies a table that is used to save the model.
required parametermodelOut	—	Specifies a table that is used to save model structure and weights.

Parameter Descriptions

* actorModel={{nnParams-1} <, {nnParams-2}, ...>}

Specifies the properties of the layers of the neural net used to build the actor network.

The nnParams value can be one or more of the following:

act="IDENTITY" | "RELU" | "SIGMOID" | "TANH"

Activation function.

Default	RELU

n=integer

Number of neurons.

Range	1–100000

size=integer

Size (width and height) of kernel (for convolution layer) or window (for pooling layer).

Range	1–1000

stride=integer

Stride of kernel (for convolution layer) or window (for pooling layer).

Default	1
Range	1–10000

type="CONV" | "FC"

Layer type.

Default	FC

actorOptimizer={method="ADAM" | "SGD", method-specific-parameters}

The optimization method for the actor model: SGD or ADAM.

Alias	actorOptimization

The value that you specify for the method parameter determines the other parameters that apply.

checkPointBestOut={casouttable}

Specifies a table that to save the best observed model.

For more information about specifying the checkPointBestOut parameter, see the common casouttable parameter.

checkPointFreq=integer

Specifies how often to save training parameters.

Range	1–100000000

checkPointIn={castable}

Specifies a table that includes a saved model for warm starting.

For more information about specifying the checkPointIn parameter, see the common castable parameter.

checkPointOut={casouttable}

Specifies a table that is used to save the model.

For more information about specifying the checkPointOut parameter, see the common casouttable parameter.

criticModel={{nnParams-1} <, {nnParams-2}, ...>}

Specifies the properties of the layers of the neural net used to build the critic network. If this value is left unspecified, the value from actorModel is copied.

The nnParams value can be one or more of the following:

act="IDENTITY" | "RELU" | "SIGMOID" | "TANH"

Activation function.

Default	RELU

n=integer

Number of neurons.

Range	1–100000

size=integer

Size (width and height) of kernel (for convolution layer) or window (for pooling layer).

Range	1–1000

stride=integer

Stride of kernel (for convolution layer) or window (for pooling layer).

Default	1
Range	1–10000

type="CONV" | "FC"

Layer type.

Default	FC

criticOptimizer={method="ADAM" | "SGD", method-specific-parameters}

The optimization method for the critic model: SGD or ADAM. If this value is left unspecified, the value from actorOptimizer is copied.

Alias	criticOptimization

The value that you specify for the method parameter determines the other parameters that apply.

* environment={type="BUILTIN" | "REMOTE", type-specific-parameters}

The reinforcement learning environment.

Alias	env

The value that you specify for the type parameter determines the other parameters that apply.

gamma=double

Discount factor used to compute long term value.

Default	0.99
Range	0–1

gpu={gpuOptions}

Specifies GPU related options.

The gpuOptions value can be one or more of the following:

deterministic=true | false

Specifies that the action will use deterministic GPU functions (may be slower).

Default	false

device=integer

Specifies a specific GPU device id.

Minimum value	0

enable=true | false

Specifies that action will use a GPU to perform model calculations.

Default	false

maxNormOfGradient=double

Adjust the magnitude of the gradient so that it is less than this value.

Range	1–100000000

* modelOut={casouttable}

Specifies a table that is used to save model structure and weights.

For more information about specifying the modelOut parameter, see the common casouttable parameter.

nThreads=integer

Number of CPU threads to use.

Range	0–1024

numEpisodes=integer

Number of episodes to train before terminating.

Default	10
Range	1–100000000

numTestEpisodes=integer

Specifies number of episodes to run when running test.

Default	100
Range	1–100000000

* pgAlgorithm={type="ACTORCRITIC" | "REINFORCE" | "REINFORCERTG", type-specific-parameters}

The policy gradient algorithm type, REINFORCE, REINFORCERTG, ActorCritic, A2C

Alias	pgAlg

The value that you specify for the type parameter determines the other parameters that apply.

seed=double

Seed for random number generator.

Minimum value	0

testInterval=integer

Specifies how often to run test.

Default	100
Range	1–100000000

Parameters for method="ADAM"

beta1=double

The exponential decay rate for the first moment estimates.

Default	0.9
Range	0–1

beta2=double

The exponential decay rate for the second moment estimates.

Default	0.999
Range	0–1

epsilon=double

The value to add into the second moment of the gradient in the denominator to improve numerical stability.

Alias	eps
Default	1E-08
Range	0–1

learningRate=double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

miniBatchSize=integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

regL2=double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

useAMSGrad=true | false

Use AMSGrad variant of ADAM. See paper: On the Convergence of Adam and Beyond.

Default	false

Parameters for method="SGD"

damping=double

Damping for momentum.

Default	0
Range	0–1

learningRate=double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

miniBatchSize=integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

momentum=double

Specifies the value for momentum for SGD.

Default	0
Range	0–1

regL2=double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

useNesterov=true | false

Enables Nesterov momentum.

Default	false

Parameters for method="ADAM"

beta1=double

The exponential decay rate for the first moment estimates.

Default	0.9
Range	0–1

beta2=double

The exponential decay rate for the second moment estimates.

Default	0.999
Range	0–1

epsilon=double

The value to add into the second moment of the gradient in the denominator to improve numerical stability.

Alias	eps
Default	1E-08
Range	0–1

learningRate=double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

miniBatchSize=integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

regL2=double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

useAMSGrad=true | false

Use AMSGrad variant of ADAM. See paper: On the Convergence of Adam and Beyond.

Default	false

Parameters for method="SGD"

damping=double

Damping for momentum.

Default	0
Range	0–1

learningRate=double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

miniBatchSize=integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

momentum=double

Specifies the value for momentum for SGD.

Default	0
Range	0–1

regL2=double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

useNesterov=true | false

Enables Nesterov momentum.

Default	false

Parameters for type="BUILTIN"

* name="string"

Parameters for type="REMOTE"

imageStack=integer

Specifies number of consecutive observations to be stacked as an input observation.

Default	1
Range	1–10000

* name="string"

The environment name used in RL.

render=true | false

Whether render the environment display or not.

Default	false

renderFreq=integer

Specifies how frequently the environment is rendered.

Alias	renderFrequency
Default	1
Range	1–10000000

renderSleep=double

Time to sleep (in seconds) between rendering frames.

Default	0

seed=integer

Seed of the environment.

Default	0

* url="string"

The url:port used for in RL training.

Default	"127.0.0.1:8080"

wrapper={"string-1" <, "string-2", ...>}

A list of strings that specifies the wrappers that will wrap the environment. The wrappers are used to preprocess the environment observations and rewards.

Parameters for type="ACTORCRITIC"

entropyCoef=double

Specifies the value for the entropy coefficient in policy gradient algorithms.

Default	0.01
Range	0–1

GAELambda=double

Specifies the value of the lambda in generalized advantage estimation (GAE).

Default	0.9
Range	0–1

nRollouts=integer

Specifies the number of rollouts to generate data.

Default	1
Range	1–1024

useGAELoss=true | false

Use GAE Loss.

Default	false

Parameters for type="REINFORCE"

entropyCoef=double

Specifies the value for the entropy coefficient in policy gradient algorithms.

Default	0.01
Range	0–1

nRollouts=integer

Specifies the number of rollouts to generate data.

Default	1
Range	1–1024

Parameters for type="REINFORCERTG"

entropyCoef=double

Specifies the value for the entropy coefficient in policy gradient algorithms.

Default	0.01
Range	0–1

nRollouts=integer

Specifies the number of rollouts to generate data.

Default	1
Range	1–1024

rlTrainPolicyGradient Action

Do reinforcement learning training using Policy Gradient algorithm.

Alias	rlTrainPG

Python Syntax

results=s.reinforcementLearn.rlTrainPolicyGradient(

=[{

"act":"IDENTITY" | "RELU" | "SIGMOID" | "TANH",

"n":integer,

"size":integer,

"stride":integer,

"type":"CONV" | "FC"

}<, {...}>],

actorOptimizer={"method":"ADAM" | "SGD", method-specific-parameters},

"caslib":"string",

"compress":True | False,

"indexVars":["variable-name-1" <, "variable-name-2", ...>],

"label":"string",

"lifetime":64-bit-integer,

"maxMemSize":64-bit-integer,

"memoryFormat":"DVR" | "INHERIT" | "STANDARD",

"name":"table-name",

"onDemand":True | False,

"promote":True | False,

"replace":True | False,

"replication":integer,

"threadBlockSize":64-bit-integer,

"timeStamp":"string",

"where":["string-1" <, "string-2", ...>]

checkPointFreq=integer,

"caslib":"string",

"computedOnDemand":True | False,

"computedVars"

:[{

"format":"string",

"formattedLength":integer,

"label":"string",

"name":"variable-name",

"nfd":integer,

"nfl":integer

}<, {...}>],

"computedVarsProgram":"string",

"dataSourceOptions":{"key-1":{any-list-or-data-type-1} <, "key-2":{any-list-or-data-type-2}, ...>},

"groupBy"

:[{

"format":"string",

"formattedLength":integer,

"label":"string",

"name":"variable-name",

"nfd":integer,

"nfl":integer

}<, {...}>],

"groupByMode":"NOSORT" | "REDISTRIBUTE",

"importOptions":{"fileType":"ANY" | "AUDIO" | "AUTO" | "BASESAS" | "CSV" | "DELIMITED" | "DOCUMENT" | "DTA" | "ESP" | "EXCEL" | "FMT" | "HDAT" | "IMAGE" | "JMP" | "LASR" | "PARQUET" | "SOUND" | "SPSS" | "VIDEO" | "XLS", fileType-specific-parameters},

"name":"table-name",

"onDemand":True | False,

"orderBy"

:[{

"format":"string",

"formattedLength":integer,

"label":"string",

"name":"variable-name",

"nfd":integer,

"nfl":integer

}<, {...}>],

"singlePass":True | False,

"vars"

:[{

"format":"string",

"formattedLength":integer,

"label":"string",

"name":"variable-name",

"nfd":integer,

"nfl":integer

}<, {...}>],

"where":"where-expression",

"whereTable"

"casLib":"string"

"dataSourceOptions":{adls_noreq-parameters | bigquery-parameters | cas_noreq-parameters | clouddex-parameters | db2-parameters | dnfs-parameters | esp-parameters | fedsvr-parameters | gcs_noreq-parameters | greenplum-parameters | hadoop-parameters | hana-parameters | hdfs-parameters | impala-parameters | jdbc-parameters | mongodb-parameters | mysql-parameters | netezza-parameters | odbc-parameters | oracle-parameters | path-parameters | postgres-parameters | redshift-parameters | s3-parameters | sapiq-parameters | sforce-parameters | singlestore_standard-parameters | snowflake-parameters | spark-parameters | spde-parameters | sqlserver-parameters | ss_noreq-parameters | teradata-parameters | vertica-parameters | yellowbrick-parameters}

"name":"table-name"

"vars"

:[{

"format":"string",

"formattedLength":integer,

"label":"string",

"name":"variable-name",

"nfd":integer,

"nfl":integer

}<, {...}>]

"where":"where-expression"

}

"caslib":"string",

"compress":True | False,

"indexVars":["variable-name-1" <, "variable-name-2", ...>],

"label":"string",

"lifetime":64-bit-integer,

"maxMemSize":64-bit-integer,

"memoryFormat":"DVR" | "INHERIT" | "STANDARD",

"name":"table-name",

"onDemand":True | False,

"promote":True | False,

"replace":True | False,

"replication":integer,

"threadBlockSize":64-bit-integer,

"timeStamp":"string",

"where":["string-1" <, "string-2", ...>]

=[{

"act":"IDENTITY" | "RELU" | "SIGMOID" | "TANH",

"n":integer,

"size":integer,

"stride":integer,

"type":"CONV" | "FC"

}<, {...}>],

criticOptimizer={"method":"ADAM" | "SGD", method-specific-parameters},

environment={"type":"BUILTIN" | "REMOTE", type-specific-parameters},

gamma=double,

"deterministic":True | False,

"device":integer,

"enable":True | False

maxNormOfGradient=double,

"caslib":"string",

"compress":True | False,

"indexVars":["variable-name-1" <, "variable-name-2", ...>],

"label":"string",

"lifetime":64-bit-integer,

"maxMemSize":64-bit-integer,

"memoryFormat":"DVR" | "INHERIT" | "STANDARD",

"name":"table-name",

"onDemand":True | False,

"promote":True | False,

"replace":True | False,

"replication":integer,

"threadBlockSize":64-bit-integer,

"timeStamp":"string",

"where":["string-1" <, "string-2", ...>]

nThreads=integer,

numEpisodes=integer,

numTestEpisodes=integer,

pgAlgorithm={"type":"ACTORCRITIC" | "REINFORCE" | "REINFORCERTG", type-specific-parameters},

seed=double,

testInterval=integer

)

indicates a required parameter

Summary: Input and Output Tables

If a row includes a subparameter, you can specify the name, caslib, and so on in the subparameter. Otherwise, you can specify the name, caslib, and so on in the parameter.

Parameters for Reading Input Tables
Parameter	Subparameter	Description
checkPointIn	—	Specifies a table that includes a saved model for warm starting.

Parameters for Creating Output Tables
Parameter	Subparameter	Description
checkPointBestOut	—	Specifies a table that to save the best observed model.
checkPointOut	—	Specifies a table that is used to save the model.
required parametermodelOut	—	Specifies a table that is used to save model structure and weights.

Parameter Descriptions

* actorModel=[{nnParams-1} <, {nnParams-2}, ...>]

Specifies the properties of the layers of the neural net used to build the actor network.

The nnParams value can be one or more of the following:

"act":"IDENTITY" | "RELU" | "SIGMOID" | "TANH"

Activation function.

Default	RELU

"n":integer

Number of neurons.

Range	1–100000

"size":integer

Size (width and height) of kernel (for convolution layer) or window (for pooling layer).

Range	1–1000

"stride":integer

Stride of kernel (for convolution layer) or window (for pooling layer).

Default	1
Range	1–10000

"type":"CONV" | "FC"

Layer type.

Default	FC

actorOptimizer={"method":"ADAM" | "SGD", method-specific-parameters}

The optimization method for the actor model: SGD or ADAM.

Alias	actorOptimization

The value that you specify for the method parameter determines the other parameters that apply.

checkPointBestOut={casouttable}

Specifies a table that to save the best observed model.

For more information about specifying the checkPointBestOut parameter, see the common casouttable parameter.

checkPointFreq=integer

Specifies how often to save training parameters.

Range	1–100000000

checkPointIn={castable}

Specifies a table that includes a saved model for warm starting.

For more information about specifying the checkPointIn parameter, see the common castable parameter.

checkPointOut={casouttable}

Specifies a table that is used to save the model.

For more information about specifying the checkPointOut parameter, see the common casouttable parameter.

criticModel=[{nnParams-1} <, {nnParams-2}, ...>]

Specifies the properties of the layers of the neural net used to build the critic network. If this value is left unspecified, the value from actorModel is copied.

The nnParams value can be one or more of the following:

"act":"IDENTITY" | "RELU" | "SIGMOID" | "TANH"

Activation function.

Default	RELU

"n":integer

Number of neurons.

Range	1–100000

"size":integer

Size (width and height) of kernel (for convolution layer) or window (for pooling layer).

Range	1–1000

"stride":integer

Stride of kernel (for convolution layer) or window (for pooling layer).

Default	1
Range	1–10000

"type":"CONV" | "FC"

Layer type.

Default	FC

criticOptimizer={"method":"ADAM" | "SGD", method-specific-parameters}

The optimization method for the critic model: SGD or ADAM. If this value is left unspecified, the value from actorOptimizer is copied.

Alias	criticOptimization

The value that you specify for the method parameter determines the other parameters that apply.

* environment={"type":"BUILTIN" | "REMOTE", type-specific-parameters}

The reinforcement learning environment.

Alias	env

The value that you specify for the type parameter determines the other parameters that apply.

gamma=double

Discount factor used to compute long term value.

Default	0.99
Range	0–1

gpu={gpuOptions}

Specifies GPU related options.

The gpuOptions value can be one or more of the following:

"deterministic":True | False

Specifies that the action will use deterministic GPU functions (may be slower).

Default	False

"device":integer

Specifies a specific GPU device id.

Minimum value	0

"enable":True | False

Specifies that action will use a GPU to perform model calculations.

Default	False

maxNormOfGradient=double

Adjust the magnitude of the gradient so that it is less than this value.

Range	1–100000000

* modelOut={casouttable}

Specifies a table that is used to save model structure and weights.

For more information about specifying the modelOut parameter, see the common casouttable parameter.

nThreads=integer

Number of CPU threads to use.

Range	0–1024

numEpisodes=integer

Number of episodes to train before terminating.

Default	10
Range	1–100000000

numTestEpisodes=integer

Specifies number of episodes to run when running test.

Default	100
Range	1–100000000

* pgAlgorithm={"type":"ACTORCRITIC" | "REINFORCE" | "REINFORCERTG", type-specific-parameters}

The policy gradient algorithm type, REINFORCE, REINFORCERTG, ActorCritic, A2C

Alias	pgAlg

The value that you specify for the type parameter determines the other parameters that apply.

seed=double

Seed for random number generator.

Minimum value	0

testInterval=integer

Specifies how often to run test.

Default	100
Range	1–100000000

Parameters for method="ADAM"

"beta1":double

The exponential decay rate for the first moment estimates.

Default	0.9
Range	0–1

"beta2":double

The exponential decay rate for the second moment estimates.

Default	0.999
Range	0–1

"epsilon":double

The value to add into the second moment of the gradient in the denominator to improve numerical stability.

Alias	eps
Default	1E-08
Range	0–1

"learningRate":double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

"miniBatchSize":integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

"regL2":double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

"useAMSGrad":True | False

Use AMSGrad variant of ADAM. See paper: On the Convergence of Adam and Beyond.

Default	False

Parameters for method="SGD"

"damping":double

Damping for momentum.

Default	0
Range	0–1

"learningRate":double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

"miniBatchSize":integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

"momentum":double

Specifies the value for momentum for SGD.

Default	0
Range	0–1

"regL2":double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

"useNesterov":True | False

Enables Nesterov momentum.

Default	False

Parameters for method="ADAM"

"beta1":double

The exponential decay rate for the first moment estimates.

Default	0.9
Range	0–1

"beta2":double

The exponential decay rate for the second moment estimates.

Default	0.999
Range	0–1

"epsilon":double

The value to add into the second moment of the gradient in the denominator to improve numerical stability.

Alias	eps
Default	1E-08
Range	0–1

"learningRate":double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

"miniBatchSize":integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

"regL2":double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

"useAMSGrad":True | False

Use AMSGrad variant of ADAM. See paper: On the Convergence of Adam and Beyond.

Default	False

Parameters for method="SGD"

"damping":double

Damping for momentum.

Default	0
Range	0–1

"learningRate":double

Specifies the learning rate parameter for SGD.

Alias	lr
Default	0.001
Range	0–1

"miniBatchSize":integer

Specifies minibatch size to be used when training the Q model.

Default	64
Range	1–100000

"momentum":double

Specifies the value for momentum for SGD.

Default	0
Range	0–1

"regL2":double

Specifies the L2 regularization parameter, the value must be nonnegative.

Alias	weightDecay
Default	0
Range	0–1

"useNesterov":True | False

Enables Nesterov momentum.

Default	False

Parameters for type="BUILTIN"

* "name":"string"

Parameters for type="REMOTE"

"imageStack":integer

Specifies number of consecutive observations to be stacked as an input observation.

Default	1
Range	1–10000

* "name":"string"

The environment name used in RL.

"render":True | False

Whether render the environment display or not.

Default	False

"renderFreq":integer

Specifies how frequently the environment is rendered.

Alias	renderFrequency
Default	1
Range	1–10000000

"renderSleep":double

Time to sleep (in seconds) between rendering frames.

Default	0

"seed":integer

Seed of the environment.

Default	0

* "url":"string"

The url:port used for in RL training.

Default	"127.0.0.1:8080"

"wrapper":["string-1" <, "string-2", ...>]

A list of strings that specifies the wrappers that will wrap the environment. The wrappers are used to preprocess the environment observations and rewards.

Parameters for type="ACTORCRITIC"

"entropyCoef":double

Specifies the value for the entropy coefficient in policy gradient algorithms.

Default	0.01
Range	0–1

"GAELambda":double

Specifies the value of the lambda in generalized advantage estimation (GAE).

Default	0.9
Range	0–1

"nRollouts":integer

Specifies the number of rollouts to generate data.

Default	1
Range	1–1024

"useGAELoss":True | False

Use GAE Loss.

Default	False

Parameters for type="REINFORCE"

"entropyCoef":double

Specifies the value for the entropy coefficient in policy gradient algorithms.

Default	0.01
Range	0–1

"nRollouts":integer

Specifies the number of rollouts to generate data.

Default	1
Range	1–1024

Parameters for type="REINFORCERTG"

"entropyCoef":double

Specifies the value for the entropy coefficient in policy gradient algorithms.

Default	0.01
Range	0–1

"nRollouts":integer

Specifies the number of rollouts to generate data.

Default	1
Range	1–1024

rlTrainPolicyGradient Action

Do reinforcement learning training using Policy Gradient algorithm.

Alias	rlTrainPG

R Syntax

results <– cas.reinforcementLearn.rlTrainPolicyGradient(s,

=list( list(

act="IDENTITY" | "RELU" | "SIGMOID" | "TANH",

n=integer,

size=integer,

stride=integer,

type="CONV" | "FC"

) <, list(...)>),

actorOptimizer=list(method="ADAM" | "SGD", method-specific-parameters),

=list(

caslib="string",

compress=TRUE | FALSE,

indexVars=list("variable-name-1" <, "variable-name-2", ...>),

label="string",

lifetime=64-bit-integer,

maxMemSize=64-bit-integer,

memoryFormat="DVR" | "INHERIT" | "STANDARD",

name="table-name",

onDemand=TRUE | FALSE,

promote=TRUE | FALSE,

replace=TRUE | FALSE,

replication=integer,

threadBlockSize=64-bit-integer,

timeStamp="string",

where=list("string-1" <, "string-2", ...>)

checkPointFreq=integer,

=list(

caslib="string",

computedOnDemand=TRUE | FALSE,

computedVars

=list( list(

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

) <, list(...)>),

computedVarsProgram="string",

dataSourceOptions=list(key-1=list(any-list-or-data-type-1) <, key-2=list(any-list-or-data-type-2), ...>),

groupBy

=list( list(

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

) <, list(...)>),

groupByMode="NOSORT" | "REDISTRIBUTE",

name="table-name",

onDemand=TRUE | FALSE,

orderBy

=list( list(

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

) <, list(...)>),

singlePass=TRUE | FALSE,

vars

=list( list(

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

) <, list(...)>),

where="where-expression",

whereTable

=list(

casLib="string"

name="table-name"

vars

=list( list(

format="string",

formattedLength=integer,

label="string",

name="variable-name",

nfd=integer,

nfl=integer

) <, list(...)>)

where="where-expression"

)

=list(

caslib="string",

compress=TRUE | FALSE,

indexVars=list("variable-name-1" <, "variable-name-2", ...>),

label="string",

lifetime=64-bit-integer,

maxMemSize=64-bit-integer,

memoryFormat="DVR" | "INHERIT" | "STANDARD",

name="table-name",

onDemand=TRUE | FALSE,

promote=TRUE | FALSE,

replace=TRUE | FALSE,

replication=integer,

threadBlockSize=64-bit-integer,

timeStamp="string",

where=list("string-1" <, "string-2", ...>)

=list( list(

act="IDENTITY" | "RELU" | "SIGMOID" | "TANH",

n=integer,

size=integer,

stride=integer,

type="CONV" | "FC"

) <, list(...)>),

criticOptimizer=list(method="ADAM" | "SGD", method-specific-parameters),

environment=list(type="BUILTIN" | "REMOTE", type-specific-parameters),

gamma=double,

=list(

deterministic=TRUE | FALSE,

device=integer,

enable=TRUE | FALSE

maxNormOfGradient=double,