Reinforcement Learning Grid World multi-figures

Hello,

I did my own version of Grid World with my own obstacles (see Code below).

My Question ist: How can I simulate the trained agent in the enviroment in multiple figures?

I am using:

plot(env)
env.Model.Viewer.ShowTrace = true;
env.Model.Viewer.clearTrace;
sim(agent,env)

And getting one variation. I tried using:

for i=1:3
    figure(i)
    plot(env)
    env.Model.Viewer.ShowTrace = true;
    env.Model.Viewer.clearTrace;
    sim(agent,env)
end

But it didn't work as planned.

Here my code for that. For some reason, I am getting spikes in the reward plot, although this already converged. I tried to tune some variables like LearnRate, Epsilon and DiscountFactor, but this is the best result I am getting of that:

GitterWelt = createGridWorld(7,7);
GitterWelt.CurrentState = '[1,1]';
GitterWelt.ObstacleStates = ["[5,3]";"[5,4]";"[5,5]";"[4,5]";"[3,5]"];
GitterWelt.TerminalStates = '[6,6]';

updateStateTranstionForObstacles(GitterWelt)

nS = numel(GitterWelt.States);
nA = numel(GitterWelt.Actions);
GitterWelt.R = -1*ones(nS,nS,nA);
GitterWelt.R(:,state2idx(GitterWelt,GitterWelt.TerminalStates),:) = 10;

env = rlMDPEnv(GitterWelt);

qTable = rlTable(getObservationInfo(env), getActionInfo(env));

qRep = rlQValueRepresentation(qTable, Obs_Info, Act_Info);

%% All trivial until here

qRep.Options.LearnRate = 0.2; % Alpha: This was in the example 1, but it doesn't make sense
Ag_Opts = rlQAgentOptions;

Ag_Opts.DiscountFactor = 0.9; % Gamma
Ag_Opts.EpsilonGreedyExploration.Epsilon = 0.02;
agent = rlQAgent(qRep,Ag_Opts);

Train_Opts = rlTrainingOptions;

Train_Opts.MaxEpisodes = 1000;
Train_Opts.MaxStepsPerEpisode = 40;
Train_Opts.StopTrainingCriteria = "AverageReward";
Train_Opts.StopTrainingValue = 10;
Train_Opts.Verbose = 1;
trainOpts.ScoreAveragingWindowLength = 30;
Train_Opts.Plots = "training-progress";


Train_Info = train(agent,env,Train_Opts);

Answer 1

[Emmanouil Tzorakoleftherakis]

Hello,

I wouldn't worry about the spikes as long as the average reward has converged. Could be the agent exploring something.

For your plotting question, the plot function for the gridworld environments has been set up with a listener callback so that it can be updated on the fly every time you call step. This means that you can only have one plot per grid world environment.

A quick workaround would be to create separate environment objects for the same grid world you created and call plot for each one. So:

function env = MyGridWorld

    GitterWelt = createGridWorld(7,7);
    GitterWelt.CurrentState = '[1,1]';
    GitterWelt.ObstacleStates = ["[5,3]";"[5,4]";"[5,5]";"[4,5]";"[3,5]"];
    GitterWelt.TerminalStates = '[6,6]';
    updateStateTranstionForObstacles(GitterWelt)
    nS = numel(GitterWelt.States);
    nA = numel(GitterWelt.Actions);
    GitterWelt.R = -1*ones(nS,nS,nA);
    GitterWelt.R(:,state2idx(GitterWelt,GitterWelt.TerminalStates),:) = 10;

    env = rlMDPEnv(GitterWelt);
end

and then

env1 = MyGridWorld;
env2 = MyGridWorld;
plot(env1)
plot(env2)

Hope that helps

[Reinforcement Learning]

Thank you very much! That worked

I actually was thinking in having them all in one figure, so I tried:

subplot(3,1,1)
env1 = MyGridWorld;
plot(env1)
env1.Model.Viewer.ShowTrace = true;
env1.Model.Viewer.clearTrace;
sim(agent,env1)

subplot(3,1,2)
env2 = MyGridWorld;
plot(env2)
env2.Model.Viewer.ShowTrace = true;
env2.Model.Viewer.clearTrace;
sim(agent,env2)


subplot(3,1,3)
env3 = MyGridWorld;
plot(env3)
env3.Model.Viewer.ShowTrace = true;
env3.Model.Viewer.clearTrace;
sim(agent,env3)

But this didn't work. I think this isn't going to work due to the listener callback feature, you mentioned earlier.