If you insist on continuing accusing me of something that I am not doing, don’t be surprised that Roblox isn’t used as a research platform like Unity and Godot did. You all prefer “no promotions please!” instead of strong scrutiny and transparency of the current resources!
At the same time, if I really wanted to promote my own stuff, why did I give the visibility to the other libraries which could reduce the effects of my non existent promotions? Wouldn’t I just highlight my libraries only?
Comparisons
Hello everyone! In this post, I have compiled all the machine learning libraries for Roblox, including the very old ones! This is to make sure that you are properly informed on which libraries you should choose for your projects. I made comparisons based on many things, including the developers’ credentials, quality assurance and so on.
There are currently 6 machine learning libraries for Roblox:
Below, I will show you each of the libraries’ current information, but I may add more in the future. So let’s begin!
Developer's Name, Education And Certifications
Detail
Developer’s Name
Developer’s Education
Developer’s Certifications
DataPredict
MyOriginsWorkshop
Bachelor In Computer Science In Artificial Intelligence With Honours, Masters In Artificial Intelligence (Coursework And Research)
Machine Learning Certification From Stanford Online, Deep Learning Certification From DeepLearning.ai
DataPredict Neural
Same As Previous
Same As Previous
Same As Previous
Kironte’s Neural Network Library 2.0
Kironte
Unknown
Unknown
OpenML
Wyvern
Unknown
Unknown
XenTorch
ArtificialIntellect
Unknown
Unknown
EasyML
wharkk
Unknown
Unknown
Machine Learning And Deep Learning Support
Detail
Machine Learning Support
Deep Learning Support
DataPredict
Full
Partial (Feed Forward Network)
DataPredict Neural
None
Full
Kironte’s Neural Network Library 2.0
None
Partial (Feed Forward Network And Recurrent Neural Network)
OpenML
None
Partial (Feature Matrix / 2D Tensor Only)
XenTorch
None
Partial (Feed Forward Network)
EasyML
Partial (Limited Number Of Features For Feature Matrix And 3 Layers Only)
Partial (Feed Forward Network)
Reinforcement Learning And Generative Models Support
Detail
Reinforcement Learning Support
Generative Models Support
DataPredict
Partial (Model-Free Deep Reinforcement Learning, Random Network Distillation, Experience Replays)
Partial (Generative Adversarial Networks And Some Of Its Variants)
DataPredict Neural
Same As Previous
Same As Previous
Kironte’s Neural Network Library 2.0
None
None
OpenML
Partial (Deep Q Learning And Experience Replay)
None
XenTorch
None
None
EasyML
None
None
Integration
Detail
Minimum Number Of Lines To Train A Model
Tutorials
API Reference
Support
DataPredict
4
Comprehensive And High Quality
Complete And High Quality
Immediate Bug Fix. If Paid User, Then Feature And Bug Fix Requests From Them Will Likely To Be Considered First.
DataPredict Neural
7
Comprehensive And High Quality
Complete And High Quality
Same As Previous
Kironte’s Neural Network Library 2.0
6
Limited And High Quality
Complete And High Quality
Unknown
OpenML
7
Limited And Low Quality
None
Unknown
XenTorch
3
Limited And Low Quality
Limited And Low Quality
Unknown
EasyML
3
Limited And Low Quality
None
Unknown
Customizability And Scalability
Detail
Highlighted Features
Customizability
Ease Of Integration
Scalability
DataPredict
Can Do Distributed Training (Data And Model Parallelism)
High
Easy
High
DataPredict Neural
Same As Previous
High
Medium (Requires The Creation Of Individual Layers And A Separate Loss Function.)
High
Kironte’s Neural Network Library 2.0
Each Neural Network Nodes Can Be Customized Individually
High
Medium (Must Define The Input And Output Names. The Inputs And Outputs Must Be Stored As Dictionary.)
Medium
OpenML
None
Medium
Hard (Nested Arrays And Functions Inside Function Parameters.)
Low
XenTorch
None
Low
Medium (Requires Calling Gradient Calculation Object For Every Activation Function Calculation Object.)
Low
EasyML
None
Low
Easy
Low
Performance And Quality Assurance
Detail
Average Training Speed For Neural Networks (in seconds)
Relative Training Time To The DataPredict Library
Quality Assurance
DataPredict
9.69946E-05
1
Unit And Integration Tests. Numerical Analysis And Gradient Checking. Research Papers And Its Source Codes. Consultation With Experts. Uses Developer’s University Lecturer Notes.
DataPredict Neural
0.00043973
4.533553009
Same As Previous
Kironte’s Neural Network Library 2.0
0.00013104
1.351000747
Unit And Integration Tests
OpenML
1.35893E-05
0.140103395
Unit And Integration Tests
XenTorch
9999999999
1.03099E+14
Unknown
EasyML
1.26872E-05
0.130802979
Unknown
Library Maturity, Cross-Compatibility And Licensing
Detail
Maturity
Cross Compatible Libraries
License
DataPredict
Stable
DataPredict Neural
Custom License - Limited Commercial Use
DataPredict Neural
Stable
DataPredict
Same As Previous
Kironte’s Neural Network Library 2.0
Stable
None
MIT License
OpenML
Stable
None
None
XenTorch
Incomplete
None
MIT License
EasyML
Incomplete
None
None
Community
Detail
Community
DataPredict
Roblox’s DevForum, Discord, Reddit, LinkedIn
DataPredict Neural
Same As Previous
Kironte’s Neural Network Library 2.0
Roblox’s DevForum
OpenML
Roblox’s DevForum
XenTorch
Roblox’s DevForum
EasyML
Roblox’s DevForum
Others
Detail
Developer Integrity
DataPredict
Transparent About Features, Limitations And Issues To The Public. Actively Maintained With A Strong Commitment To QA. If Any Promised Development Will Not Be Added, The Public Will Be Informed.
DataPredict Neural
Same As Previous
Kironte’s Neural Network Library 2.0
Claims A New Library Similar To TensorFlow Will Be Developed, But There Are No Development For The Last 3 Years.
OpenML
Claims To Add Machine Learning Models, But No Models Have Been Added For The Last 4 Months.
XenTorch
Claims To Add More Features, But There Are No New Developments For The Last 3 Years.
EasyML
Claims To Add More Models, But There Are No New Developments For The Last 4 Months.
This section is very biased… you only know about your own skillsets and not the others. It’s an unfair comparison this way!
Anyways, just at a glance at all the resources, their interfaces, use cases, and performance, DataPredict seems to be the best. None of these have a tutorial I could actually understand though.
I know it’s biased! The reason it is there so that people be more careful about other developer libraries and remind them that you can’t just use the libraries without being aware of potential incorrect implementations due to informal credentials.
I’m actually much more concerned about the other developers’ implementations of these libraries, as this could result reputation damage for machine learning for Roblox as a whole. People from the academia might want to implement experiments in Roblox, and I’m very concerned about getting wrong results due to potential incorrect implementations. In fact, I already seen some problematic ones from OpenML when it comes to convolution layers and pooling layers.
Non, because I don’t like artifical intelligence and I don’t understand it much. Would be nice if you can make maybe a guide for how it actually works in a way that you are explaining it to a 9 year.
@MYOriginsWorkshop and @Wyv_n you two should consider arguing in the DM’s as if you guys argue publically then it’s pretty bad for your reputation.
Non, because I don’t like artifical intelligence and I don’t understand it much. Would be nice if you can make maybe a guide for how it actually works in a way that you are explaining it to a 9 year.
Artificial Intelligence is really broad term though. Although it can cover machine learning and deep learning, it is mainly concerned about mimicking human intelligence. You can actually count pathfinding as AI actually.
@MYOriginsWorkshop and @Wyv_n you two should consider arguing in the DM’s as if you guys argue publically then it’s pretty bad for your reputation.
Meh. He decided to bring it to the public. I thought it would be much more appropriate for me to move it to DM even though he started first. But I thought it would a great idea to leave it public to increase the “reply” count and so more people will join us in and see this post more.
I know about it’s literal definition yet don’t know how the machine actually learns. I do understand using fitness scoring to see which machine did better yet actually making it “learn” is a challenge.
Yes, I do know about that. Almost every algorithm can be considered AI. I have also developed some pathfinding algorithms on roblox.
That’s kind of down bad of you ngl… But whatever, I don’t want to argue about this.
I know about it’s literal definition yet don’t know how the machine actually learns. I do understand using fitness scoring to see which machine did better yet actually making it “learn” is a challenge.
I guess let’s explain it in more human term then.
So let’s say you have a cannon that shoots a ball upwards and lands at the exact same spot. From here, you want to learn to catch it perfectly. When the cannon shoots the ball, you will try to “predict” where the ball will land. Unfortunately, for your first time, you didn’t catch it because your hand is way too far away from the ball. The distance between the ball and your hand is what they call an “error”. Now you learn to have your hand to be closer to that location by large amount. So large “error”.
The next time the ball shoot at the same location, your hand nearly catches it, and hence lower “error”. So you move slightly closer. You will repeat these steps again until one of these happens:
a) You can’t improve your location, but you still able to catch it, just not perfectly.
b) You finally able to catch the ball.
Then, you use that location as your “prediction” where the ball lands.
That’s just how the machine learning and deep learning works…
Your making it sound like roblox is a platform where only developers with credentials should be able to create. I hope you remember this platform is target towards kids, and for them to learn, not to be perfect.
No it won’t this whole platform is meant so people learn, yeah someones going to make a mistakes but thats what the devforum is for, to help each other.
This whole post just seems like a way to bump your library, and there for I actually see no reason for this post… I mean you yourself said the post was biased so there really isn’t an accurate comparison made in this post…
Your making it sound like roblox is a platform where only developers with credentials should be able to create. I hope you remember this platform is target towards kids, and for them to learn, not to be perfect.
I didn’t say the roblox games themselves should be developed by people who has credential, did I? I was specifically referring to the machine learning and deep learning libraries which would be used by researchers, which could lead problematic results if the library developers have no such credential in place. Do you want researchers to use Roblox as research grounds so that we have better resources? If yes, then these kind of stuff need to be held at high standard.
No it won’t this whole platform is meant so people learn, yeah someones going to make a mistakes but thats what the devforum is for, to help each other.
Oh, people are learning! Learning the existence of other people’s libraries if they aren’t aware of them yet.
This whole post just seems like a way to bump your library, and there for I actually see no reason for this post… I mean you yourself said the post was biased so there really isn’t an accurate comparison made in this post…
Then feel free to ignore it then. Let it slide into the abyss.
Almost same thing here. Although, this is getting into another territory called “reinforcement learning”.
obstacle avoiding
Let’s say we’re playing a game where we can only dash forward, left and right. And we have three tracks, where the trains can move towards us. We have left, middle and right. You have to repeat the game when you get hit by the train. The longer we survive, the better high score (rewards) we get. Crash? Then hurt (heavy punishment).
Now let’s say we are spawning in the middle track, and a train is coming right towards us at the middle track as well. We have two issues here:
Determining the distance between us and the train before we can switch tracks.
Determining the direction which to move.
Since this is your first time in your game, you can expect that you don’t know the train’s speed and what could get you killed. This is what happens when we create an untrained model a.k.a. “We don’t know anything about this game” phase.
First time you play, you didn’t react fast enough when the train crashes you. You know that the crashing hurts you, so you try to avoid it. However, because you aren’t used to the train’s speed, you crash against it every time due to low reaction time. The more you crash, however, you start to be able to “predict” when to move since you can now determine the speed.
Eventually, you avoid the first train. Unfortunately for you, that’s not where the game ends. So few more trains incoming! Now let’s say we have two trains are coming now. This gives us an issue:
Determine the sequence to avoid the two train.
Determine the timing when to avoid for each train.
At this point, you crash against them multiple times, but somehow, you started to learn a “pattern” to avoid particular situations. Sometimes you avoid, sometimes you crash. Eventually you get a point where the number of crashes begin to decrease and you get better high score.
For this part, it’s either they learn by chance or trial-and-error.
For chance, they might have gotten lucky that they did not crash into a train (i.e. randomly going left or right"). So, the model “learns” that particular sequence of movements leads to rewards.
For trial and error part, it might keep on crashing multiple times, leading to the model to less likely to choose that particular action. Then when the probability to choose that action is low enough, the model starts considering to choose other actions, which could potentially lead to point 1.
Another one went thinking that I am doing advertisement…
Here’s a short message from frames:
Your making it sound like roblox is a platform where only developers with credentials should be able to create. I hope you remember this platform is target towards kids, and for them to learn, not to be perfect.
My message:
I didn’t say the roblox games themselves should be developed by people who has credential, did I? I was specifically referring to the machine learning and deep learning libraries which would be used by researchers, which could lead problematic results if the library developers have no such credential in place. Do you want researchers to use Roblox as research grounds so that we have better resources? If yes, then these kind of stuff need to be held at high standard.
