Deep Learning - TensorFlow

Hi, I tried the machine learning puzzle (which by the way is an awesome idea, thanks a lot !)
I was just wondering : Will it be possible to have more time to provide an output ? For the advanced technique for example (the Multilayer Convolutional Network), it takes quite a bit of time to train the model.


How much time would you need?

2 weeks ? :smile:

1 Like

Well, for the advanced technique, the timeout occurs after 38 loops of training (on 20000), so I guess it would need ~550x more time ^^;

Yep I also find that the timeout occurs between 30 and 40 loops approximately.
I ran this on my laptop and 500 loops are needed to reach something like 95% of correct answers. I began to reach the 97% around 1000 loops (a bit less) so I guess it would need too much time to train.

That’s our main current issue with Machine Learning: if you want to implement a more advanced technique, you need at least minutes (if not hours/days) of computing. One solution could be to allow users to upload a data file that could contained your learning data. This way you could train your model outside of CodinGame for really time consuming trainings. What do you think?

If you have any idea of games/puzzles that could be interesting with Machine Learning but not too time consuming we would be happy to hear from you.

Yeah sure, it could be a good idea for time consuming training. How will we load our model on the site ? With something like Pickle ?
But it would mean that everyone has the possibility to train a better model with more data if they find some and so it may be a problem if there is a score-based challenged or similar (unless it’s part of the rules :p)

Yep if I find some good ideas for ML puzzles I will propose them (via the form right ?)

Thanks for the fast reply by the way

Pour Maxime : il existe peu de cas de jeu et ou de puzzle dans lequel le machine learning peut remplacer (de façon au moins Ă©quivalente) la force brute ou l’algorithmique standard.
Surtout si vous limitez les diffĂ©rentes phases (notamment d’apprentissage) Ă  moins de quelques secondes.
NĂ©anmoins, je pourrais citer ; parce que ces domaines des jeux ne sont pas encore couverts par Codingame et pourrait effectivement ĂȘtre pertinent pour du machine learning

  • Dans la thĂ©orie des jeux, il existe tout un pĂ©rimĂštre sur les jeux Ă  somme non nulle ; lĂ -dessus, il y a ENORMEMENT de matiĂšre de jeux Ă  faire et trĂšs peu explorĂ©. Y compris en mode compĂ©titif ; peut-ĂȘtre regarder de ce cĂŽtĂ©.

  • trouver des jeux ou l’apprentissage est nĂ©cessaire parce que l’algorithmique est quasi impossible et la force brute inutile ; en clair ou les possibilitĂ©s sont nombreuses. Avec certains challenges c’était presque possible (beaucoup de pions, notamment les platinums rift). Il faut aller plus loin en proposant un jeu combinant beaucoup de piĂšces, diffĂ©rentes dans leur comportement, un aspect tactique correct et un aspect stratĂ©gique plus important que l’'aspect tactique ; c’est le domaine des RTS 
 d’ailleurs DeepMind est en train d’aller sur ce terrain ; vous savez ce qu’il vous reste Ă  faire ; un moteur de jeu RTS ou des bots pourraient ĂȘtre soumis 
 bonjour le travail :slightly_smiling: et peu de personne seront intĂ©ressĂ©s car le ticket d’entrĂ©e est lourd ! Bref, Ă©normĂ©ment de travail pour peu de joueurs, est-ce intĂ©ressant ? sauf Ă  standardiser une plateforme de bots RTS (standard non existant).

Par contre, des challenges sur des jeux Ă  somme non nulle, c’est une bonne idĂ©e (je sors du sujet Machine Learning), trĂšs facile Ă  implĂ©menter (quasi pas de graphismes) et des stratĂ©gies possibles trĂšs diverses (coopĂ©ratif ou pas, 

1 Like

Maxime, i think that it would be interesting to implement the QWOP game ( as a AI game for machine learning because its a game where it would be very hard to implement a standard algorithm. It could be implemented as a competition in 2 way: 1) a race where 2-4 runner bots need to finish a distance first or 2) a endurance challenge where bots need to run the biggest distance before falling

Hope you like the suggestion,
Samuel Croteau


I just found this video as a example of AI that play the game, this look so fun to program and its so funny to watch the bot strategy that they learn by themselves xD

Honestly I wish my job was to build challenge like that, you guys are lucky to program nice competition like you are doing now :stuck_out_tongue:


1 Like

In the description, it is written that “evolution took a total time of about 20 hours”. That’s sad because that implies we can’t do the training on the website directly. But it’s a great idea, I’d love to see that on CodinGame :D.

1 Like

Apologies if I am being stupid here and this is part of the challenge/we are supposed to run code locally, but I seem unable to import the tensorflow library and/or input_data on here, even though this is attempted in the code provided in the puzzle. It is throwing an error complaining about lacking the ‘six’ module.

It will take more time than 2 weeks for the ouput.
(Tensorflow )

1 Like

Some top players on last challenges used neural networks or similar, all having issues with code size. Any plans to allow users to upload a datafile bigger? Not as a submit code, but a separated file that aren’t resubmitted each time.
And related to this, there is little point in having N times the same code resubmitted in history, if not changed from previous submit it can appear as a counter per submit.


Can someone post a link to the puzzle ?
I can’t find it :man_shrugging:

This puzzle was removed since some time already.

1 Like