Noughts n Crosses Winning Computer winning Condition

[C S]

Hi my students have been playing the Noughts and Crosses game in ML for kids platform, even after training the below examples, the model is not winning, what is the winning number for the computer, after how many examples it actually starts winning.

You have collected:

• 19 examples of top_left,
• 16 examples of top_middle,
• 28 examples of top_right,
• 14 examples of middle_left,
• 30 examples of middle_middle,
• 15 examples of middle_right,
• 18 examples of bottom_left,
• 20 examples of bottom_middle,
• 23 examples of bottom_right

[Dale Lane]

I shared my experience of trying this in my book - I'll include a photo of the page here:

I counted games rather than individual moves in that description. But as a winning game must contain at least 3 moves, that graph represents at least 900 moves as a minimum.

But that is just one isolated experience, and absolutely cannot be used as a guarantee of the outcome for every machine learning project. 

Beyond that, I'm afraid it is very difficult for me to diagnose the behaviour of your machine learning project based solely on the amount of training data. 

There are many reasons why it might not be doing what you expect.

For example, there might be a bug in your code that means training data is not being collected correctly. Or a bug in the code that means the wrong values are being used in classifying. Or a bug in the code that means perhaps the machine learning model is being trained with losing moves, instead of winning. 

Or it might not be related to a bug at all, and simply be a factor of the training data itself. For example, if you have many students contributing training examples who all played so differently to each other that there wasn't a clear emergent pattern in the approach that led to a win, so the machine learning model was not able to learn an approach. 

Or dozens of other reasons - without seeing the training data and code, I really can only guess blindly.

I would encourage you to use this as a learning opportunity with your students, however. Machine learning projects in the "real world" do often experience this kind of phase. Machine learning models are often described as a "black box" - diagnosing the behaviour of non-deterministic systems when they don't do what we want is a difficult and complicated task, and often involves having to review training data to try and get a "feel" for what might be happening, collecting more training examples then you had planned, and just general frustrated head-scratching. 

Kind regards

D

[C S]

Thats really a helpful insight, thank you !