I received this error in DDPG "Model input sizes must match the dimensions specified in the corresponding observation and action info specifications."

hi every body

i want to impliment LSTM layers in state path of my critic network

my code is :

obsInfo = rlNumericSpec([42 1]);

obsInfo.Name = 'observation';

actInfo = rlNumericSpec([6 1]);

actInfo.Name = 'action';

actInfo.UpperLimit =0.01*ones(6,1);

actInfo.LowerLimit =-0.01*ones(6,1);

numObs=prod(obsInfo.Dimension) ;

numAcs=prod(actInfo.Dimension) ;

numHiddenUnits=100;

statePath = [sequenceInputLayer(numObs)

lstmLayer(numHiddenUnits,'OutputMode','sequence')

fullyConnectedLayer(500,'Name','CriticStateFC1')

reluLayer('Name', 'CriticStateRelu1')

fullyConnectedLayer(450,'Name','CriticStateFC2')

reluLayer('Name', 'CriticStateRelu2')

fullyConnectedLayer(300,'Name','CriticStateFC3')

reluLayer('Name', 'CriticStateRelu3')

fullyConnectedLayer(250,'Name','CriticStateFC4')

reluLayer('Name', 'CriticStateRelu4')

fullyConnectedLayer(200,'Name','CriticStateFC5')

reluLayer('Name', 'CriticStateRelu5')

fullyConnectedLayer(100,'Name','CriticStateFC6')

reluLayer('Name', 'CriticStateRelu6')

fullyConnectedLayer(50,'Name','CriticStateFC7')

reluLayer('Name', 'CriticStateRelu7')

fullyConnectedLayer(25,'Name','CriticStateFC8')

];

actionPath = [

sequenceInputLayer(numAcs)

fullyConnectedLayer(300,'Name','CriticActionFC1')

reluLayer('Name', 'CriticActionRelu1')

fullyConnectedLayer(250,'Name','CriticActionFC2')

reluLayer('Name', 'CriticActionRelu2')

fullyConnectedLayer(200,'Name','CriticActionFC3')

reluLayer('Name', 'CriticActionRelu3')

fullyConnectedLayer(100,'Name','CriticActionFC4')

reluLayer('Name', 'CriticActionRelu4')

fullyConnectedLayer(50,'Name','CriticActionFC5')

reluLayer('Name', 'CriticActionRelu5')

fullyConnectedLayer(25,'Name','CriticActionFC6')];

commonPath = [

concatenationLayer(1,2,Name='concat')

reluLayer('Name','CriticCommonRelu')

fullyConnectedLayer(512,'Name','CriticCommon1')

reluLayer('Name','CriticCommonRelu2')

fullyConnectedLayer(300,'Name','CriticCommon2')

reluLayer('Name','CriticCommonRelu3')

fullyConnectedLayer(1,'Name','CriticOutput')];

criticOpts =rlOptimizerOptions(LearnRate=1e-04,GradientThreshold=inf,...

L2RegularizationFactor=0.0001,Optimizer="adam");

criticNetwork = layerGraph();

criticNetwork = addLayers(criticNetwork,statePath);

criticNetwork = addLayers(criticNetwork,actionPath);

criticNetwork = addLayers(criticNetwork,commonPath);

criticNetwork = connectLayers(criticNetwork,'CriticStateFC8','concat/in1');

criticNetwork = connectLayers(criticNetwork,'CriticActionFC6','concat/in2');

criticNetwork = dlnetwork(criticNetwork);

analyzeNetwork(criticNetwork)

critic = rlQValueFunction(criticNetwork,obsInfo,actInfo,...

ObservationInputNames='observation',ActionInputNames='action',UseDevice="gpu");

when i run above code i received this error

"Error using rlQValueFunction

Model input sizes must match the dimensions specified in the corresponding observation

and action info specifications." please help me

Answer 1

[Emmanouil Tzorakoleftherakis]

The easiest way to discover your error yourself is to use the default agent feature have use the network architecture that's automatically generated by Reinforcement Learning Toolbox. Please take a look here. Don't forget to indicate that you want your network to use LSTM layers in the initialization options.

Per the example in the link above, after the agent is created, you can extract the actor and critic and compare the generated architecture with what you have.

Hope this helps