Environment feature vector must be in a table of table of values.
Please read the DataTypes in the API reference to see what it looks like.
1 Like
I think its:
local testData = {
{1, 90, 32},
{1, -120, -41}
}
local predictedVector = LogisticRegressionModel:predict(testData)
1 Like
Also thank you very much again. I have fixed it. Must be fatigue that caused me to missed that.
1 Like
No problem, now it’s like this:
local DQN = DataPredict.Models.QLearningNeuralNetwork.new() -- Create a new model object.
DQN:createLayers({4, 3, 2}) -- Setting up our layers.
DQN:setClassesList({"Up", "Down"}) -- Setting up our classes.
local reward = 0
wait(10)
local part = workspace:FindFirstChild("AIPart")
local oldposition = part.Position.X
DQN:setPrintReinforcementOutput(true)
while true do
local playerposition = part.Position.X
--##!!
local environmentFeatureVector = {
{playerposition,0, 0}
}
--##!!
if not (playerposition >= 50) then
if playerposition == oldposition then
reward = -1
else
reward = playerposition - oldposition
end
else
reward = 50
end
oldposition = part.Position.X
local actionLabel = DQN:reinforce(environmentFeatureVector, reward) -- Run the reinforce() function.
if actionLabel == "Up" then
local pol = part.Position
part.Position.X = pol + 1
else
local pol = part.Position
part.Position.X = pol - 1
end
if part.Position.X >= 50 then
wait(10)
end
wait(0.1)
end
unfortunately now comes the error:
12:06:41.118 ServerScriptService.MatrixL:231: Argument 1 and 2 are incompatible! (1, 3) and (5, 4) - Server - MatrixL:231
12:06:41.118 Stack Begin - Studio
12:06:41.118 Script 'ServerScriptService.MatrixL', Line 231 - function dotProduct - Studio - MatrixL:231
12:06:41.118 Script 'ServerScriptService.DataPredictLibrary.Models.NeuralNetwork', Line 368 - function forwardPropagate - Studio - NeuralNetwork:368
12:06:41.119 Script 'ServerScriptService.DataPredictLibrary.Models.NeuralNetwork', Line 960 - function predict - Studio - NeuralNetwork:960
12:06:41.119 Script 'ServerScriptService.DataPredictLibrary.Models.QLearningNeuralNetwork', Line 211 - function reinforce - Studio - QLearningNeuralNetwork:211
12:06:41.120 Script 'ServerScriptService.AI2', Line 109 - Studio - AI2:109
12:06:41.120 Stack End
Thanks for the help 
,
Sorry
And:
local testData = {
{1, 90, 32},
{1, -120, -41}
}
Doesn’t there have to be another comma?
Please excuse me that I(probably) get on your nerves so much
Two issues:
createLayers({3, 3, 2}) -- 3 inputs, without bias.
local environmentFeatureVector = {
{1, playerposition, 0, 0} -- 1 is bias
}
1 Like
Yay, thank you, it finally works, thanks to your help I made (with your module) my first AI.
Thank you!
Huh, I get this warning:
14:46:58.779 The model diverged! Please repeat the experiment again or change the argument values. - Server - NeuralNetwork:938
and a litte after this this error:
14:47:09.837 ServerScriptService.DataPredictLibrary.Models.QLearningNeuralNetwork:119: table index is nil - Server - QLearningNeuralNetwork:119
And how can you save the progress of the model so that it does not always start from 0 when you make a new test?
That warning happens if you set one of these values too high:
-
Learning rate
-
DiscountFactor
-
Reward Value
Usually these values are recommended to be between 0 and 1 for the first two. The last one should be between -1 and 1.
You can save your model using :getModelParameters() and load it using :setModelParameters(). Just look for BaseModel under Models in API Documentation.
1 Like
Thanks, my max reward is two and my max penalty is 4. But I’m only waiting 0.01 second before the next execution, that’s the problem right?
Oh, I just saw that I did something wrong and the max reward can be much higher.
I wait now 5 seconds and reworked the reward system.
But Anyway ( after 5 min of training):
16:20:22.208 The model diverged! Please repeat the experiment again or change the argument values. - Server - NeuralNetwork:938
16:20:22.209 Current Number Of Episodes: 165 Current Epsilon: 0.4990005 - Server - QLearningNeuralNetwork:239
16:20:22.209 Right - Server - AI2:181
16:20:22.209 1 - Server - AI2:156
16:20:22.744 The model diverged! Please repeat the experiment again or change the argument values. - Server - NeuralNetwork:938
16:20:22.745 Current Number Of Episodes: 166 Current Epsilon: 0.4990005 - Server - QLearningNeuralNetwork:239
16:20:22.745 Left - Server - AI2:178
16:20:22.746 -2 - Server - AI2:156
What does this mean? Current Epsilon
Reward is recommended to be between -1 and 1, but it is not necessary. As long it isn’t too high like 9999, then it would be fine. Yours seem okay.
Try adding individual layers using :addLayer() function. Might be a neural network structure issue.
Epsilon is already explained in the documentation for Q-Learning neural network.
1 Like
Added a video showcasing self-learning sword-fighting AIs that uses this library.
2 Likes
Great video, would it be possible to share the code from it?
1 Like
Take the full file instead. It would be a hassle to set things up.
1 Like
Wow,
Is there a code in the docs? 
It’s not doesn’t even give results yet. Might need some finetuning later. Once I get the result I need, then the code probably get released.
1 Like
You should give punishments if they fall down, map soon
By the way you should use an r6 rig since it will be easier in my opinion
It is R6 rig. I also have added falling and crawling punishment.
How much should they walk? Im thinking 100 studs