Replace RL type (PPO with DPPG) in a Matlab example

There is a Matlab example about coverage path planning using PPO reinforcement learning in the following link:

https://www.mathworks.com/help/reinforcement-learning/ug/train-3-agents-for-area-coverage.html

I think the environment is fine and I only need to check the parts that there is a PPO. I am trying to replace PPO with DDPG, with the following codes

opt = rlDDPGAgentOptions(...

ActorOptimizerOptions=actorOpts,...

CriticOptimizerOptions=criticOpts,....

MiniBatchSize=64,...

SampleTime=Ts,...

DiscountFactor=0.995);

agentA = rlDDPGAgent(actor(1),critic(1),opt);

agentB = rlDDPGAgent(actor(2),critic(2),opt);

agentC = rlDDPGAgent(actor(3),critic(3),opt);

but there is error: First argument must be a rlDeterministicActorRepresentation object or an observation specification created using 'rlNumericSpec' or 'rlFiniteSetSpec' objects.

Do you have any idea?

Answer 1

[Emmanouil Tzorakoleftherakis]

PPO is a stochastic agent whereas DDPG is deterministic. This means that you cannot just use actors and critics designed for PPO with DDPG and vice versa. Your best bet is to either recreate those neural nets or use the default agent feature to get an initial architecture you can iterate upon.

[ali farid]

Thanks for your input Emmanouil. Could you please guide me about neural networks recreation that can fit with deterministic DDPG. To be honest, I am a bit confused, how I can modify the neural networks. Some of actor critics are stright lines, while others are different. I appreciate if you introduce me some resources for reading.

[Emmanouil Tzorakoleftherakis]

I would start here and then take a look at the examples and how we set up the neural networks for each scenario

[ali farid]

Thanks for your guidance. I recreated those neural nets but still there is error: matlab model input sizes must match the dimensions specified in the corresponding observation and action info specification

for idx = 1:3
    % Create actor deep neural network.
    actorNetwork = [
        imageInputLayer(obsSize,Normalization="none")
        convolution2dLayer(8,16, ...
            Stride=1,Padding=1,WeightsInitializer="he")
        reluLayer
        convolution2dLayer(4,8, ...
            Stride=1,Padding="same",WeightsInitializer="he")
        reluLayer
        fullyConnectedLayer(256,WeightsInitializer="he")
        reluLayer
        fullyConnectedLayer(128,WeightsInitializer="he")
        reluLayer
        fullyConnectedLayer(64,WeightsInitializer="he")
        reluLayer
        fullyConnectedLayer(numAct)
        softmaxLayer];
    actorNetwork = dlnetwork(actorNetwork);
    
    % Create critic deep neural network.
    statePath = [...
        featureInputLayer(prod(oinfo.Dimension),Name="NetObsInLayer")
        fullyConnectedLayer(128)
        reluLayer
        fullyConnectedLayer(200,Name="sPathOut")];

    actionPath = [
        featureInputLayer(prod(ainfo.Dimension),Name="NetActInLayer")
        fullyConnectedLayer(200,Name="aPathOut",BiasLearnRateFactor=0)];



    commonPath = [
        additionLayer(2,Name="add")
        reluLayer
        fullyConnectedLayer(1,Name="CriticOutput")];

    % Create layerGraph object and add layers
    criticNetwork = layerGraph(statePath);
    criticNetwork = addLayers(criticNetwork,actionPath);
    criticNetwork = addLayers(criticNetwork,commonPath);

    % Connect paths and convert to dlnetwork object
    criticNetwork = connectLayers(criticNetwork,"sPathOut","add/in1");
    criticNetwork = connectLayers(criticNetwork,"aPathOut","add/in2");
    criticNetwork = dlnetwork(criticNetwork);

    % create actor and critic
    actor(idx) = rlDiscreteCategoricalActor(actorNetwork,oinfo,ainfo);%(actorNetwork,oinfo,ainfo); %#ok<*SAGROW> 
    critic(idx) =  rlQValueFunction(criticNetwork, ...
    oinfo,ainfo)

end