MATLAB: Implementation of Proximal Policy Optimisation

Question

I am currently trying to control the simlink homebrew environment using PPOAgent.

However, the following error occurs, and the problem continues to be unsuccessful.

How should we improve the situation?

Error: rl.representation.rlStochasticActorRepresentation (line 32)
Number of outputs for a continuous stochastic actor representation must be two times the number of actions.
Error: rlStochasticActorRepresentation (line 139)
Rep = rl.representation.rlStochasticActorRepresentation(...

my code

clear all
motion_time_constant = 0.01;
mdl = 'fivelinkrl';
open_system(mdl)
Ts = 0.05;
Tf = 20;
mdl = 'fivelinkrl';
open_system(mdl)
agentblk = [mdl '/RL Agent'];
numObs = 15;
obsInfo = rlNumericSpec([numObs 1]);
obsInfo.Name = 'observations';
numAct = 5;
actInfo = rlNumericSpec([numAct 1],'LowerLimit',-10,'UpperLimit',10);
actInfo.Name = 'Action';
% define environment
env = rlSimulinkEnv(mdl,agentblk,obsInfo,actInfo);
%createPPOAgent
criticLayerSizes = [400 300];
actorLayerSizes = [400 300];
createNetworkWeights;
criticNetwork = [imageInputLayer([numObs 1 1],'Normalization','none','Name','observations')
    fullyConnectedLayer(criticLayerSizes(1),'Name','CriticFC1', ... 
                                            'Weights',weights.criticFC1, ...
                                            'Bias',bias.criticFC1)
    reluLayer('Name','CriticRelu1')
    fullyConnectedLayer(criticLayerSizes(2),'Name','CriticFC2', ...
                                            'Weights',weights.criticFC2, ... 
                                            'Bias',bias.criticFC2)
    reluLayer('Name','CriticRelu2')
    fullyConnectedLayer(1,'Name','CriticOutput',...
                          'Weights',weights.criticOut,...
                          'Bias',bias.criticOut)];
                      
criticOpts = rlRepresentationOptions('LearnRate',1e-3);
critic = rlValueRepresentation(criticNetwork,env.getObservationInfo, ...
                          'Observation',{'observations'},criticOpts);
                      
actorNetwork = [imageInputLayer([numObs 1 1],'Normalization','none','Name','observations')
    fullyConnectedLayer(actorLayerSizes(1),'Name','ActorFC1',...
                                           'Weights',weights.actorFC1,...
                                           'Bias',bias.actorFC1)
    reluLayer('Name','ActorRelu1')
    fullyConnectedLayer(actorLayerSizes(2),'Name','ActorFC2',...
                                           'Weights',weights.actorFC2,...
                                           'Bias',bias.actorFC2)
    reluLayer('Name','ActorRelu2')
    fullyConnectedLayer(numAct,'Name','Action',...
                               'Weights',weights.actorOut,...
                               'Bias',bias.actorOut)
    softmaxLayer('Name','actionProbability')
    ];  
actorOptions = rlRepresentationOptions('LearnRate',1e-3);
%%%%  ↓error   %%%%%%%%%%%%%%%%%
actor = rlStochasticActorRepresentation(actorNetwork,obsInfo,actInfo,... 
                         'Observation',{'observations'}, actorOptions);
%%%%  ↑error   %%%%%%%%%%%%%%%%%%
opt = rlPPOAgentOptions('ExperienceHorizon',512,...
                        'ClipFactor',0.2,...
                        'EntropyLossWeight',0.02,...
                        'MiniBatchSize',64,...
                        'NumEpoch',3,...
                        'AdvantageEstimateMethod','gae',...
                        'GAEFactor',0.95,...
                        'SampleTime',0.05,...
                        'DiscountFactor',0.9995);
agent = rlPPOAgent(actor,critic,opt);  
%TrainAgent
maxEpisodes = 4000;
maxSteps = floor(Tf/Ts);
trainOpts = rlTrainingOptions(...
    'MaxEpisodes',maxEpisodes,...
    'MaxStepsPerEpisode',maxSteps,...
    'ScoreAveragingWindowLength',250,...
    'Verbose',false,...
    'Plots','training-progress',...
    'StopTrainingCriteria','EpisodeCount',...
    'StopTrainingValue',maxEpisodes,...
    'SaveAgentCriteria','EpisodeCount',...
    'SaveAgentValue',maxEpisodes);
trainingStats = train(agent,env,trainOpts);
save('agent.mat', 'agent')
Result in simulation
simOptions = rlSimulationOptions('MaxSteps',maxSteps);
experience = sim(env,agent,simOptions);

Answer 1

Hello,

It seems you want to use PPO with continuous action space. If that's the case, your actor network does not have the right architecture. With stochastic agents, the neural network should end with a path that outputs 'mean' value and another path that outputs 'variance'. In your case you seem to only have a single path. Please refer to this example here to get an idea on how to set up your actor network. Also make sure you are using 20a (PPO for continuous actions was not available in previous releases as far as I remember).

Hope that helps