Applying Machine Learning Like a Responsible Adult

In this 2015 GDC session, Havok's Ben Sunshine-Hill and The University of Pennsylvania's Aline Normoyle describe key concepts from machine learning, such as over-fitting and the bias/variance tradeoff, and the problem spaces (supervised/unsupervised, classification versus regression, etc.). In this 2015 GDC session, Havok's Ben Sunshine-Hill and The University of Pennsylvania's Aline Normoyle describe key concepts from machine learning, such as over-fitting and the bias/variance tradeoff, and the problem spaces (supervised/unsupervised, classification versus regression, etc.).

GDC talks cover a range of developmental topics including game design, programming, audio, visual arts, business management, production, online games, and much more. We post a fresh GDC video every day. Subscribe to the channel to stay on top of regular updates, and check out GDC Vault for thousands of more in-depth talks from our archives.

[MatheusMullerMeM]

Great contribution! But it's a shame that community-contributed closed caption is not enabled for this video (and others in the GDC channel).
The lack of closed captioning can make it really hard to use the videos as educational material for non-native English speakers. Even English captioning would help a lot.
I would very strongly suggest enabling this so that the community can contribute transcriptions as needed.

[civilizedmonkey]

Holy heck thanks for uploading this. I just started studying machine learning and this is very interesting to me :D

[AntonQvarfordt]

I still feel nobody's managed to give even a somewhat convincing justification for spending resources on ML in games. In 999 cases out of 1000 you want complete control of an AI's actions and you want them to be predictable and static. Mastery of a game through decoding and exploiting predictable behavior is probably in the top 3 biggest drivers for engagement in the history of gaming. 

My point is; nobody can give me an example (at least one that's applicable to a relatively large percentage of games) of why and how machine learning  would make my game more engaging.

[TrustPatches]

This is a fantastic video for understanding basics of ML in general, even outside of the context of game dev

[lordGoldemort]

Very good summary of ML. I still want to see a really good application of it in a game (yes, Black & White used Monte Carlo Tree Search, but that was an exception and it didn't work all that well). My own suspicion is that ML will not really solve the sorts of general problems users would like. I say this from experience. In 2003/4 at Microsoft, I worked with researchers to retro-fit Reinforcement Learning into a Tekken-style fighting game (only 2 players and a sparse environment, so the state space was small). It worked quite well, training the model overnight against an NPC running its state machine. Against static policies, after a few thousand training rounds, it could evolve a decent policy which could be tuned using parameters to have different fighting styles (aggressive, defensive, aikido [never hit, never get hit]...). The trouble is that a thousand rounds is a long time, and against a dynamic opponent which switches strategy in real time, it cannot hope to compete. This is why ML is interesting, but currently not the Golden Hammer that it is portrayed as. Note that the Q function representation in this game was a perceptron - a very basic neural network (tables don't work well). In modern applications, the networks are massive, often implemented in TensorFlow, which doesn't fit in with a game development environment in terms of memory usage, or performance.

I think, contrary perhaps to some, that ML will prove useful in non-gameplay contexts. For example tuning difficulty levels based on player performance to generate new procedural content that challenges players without being too easy (they walk away bored) or too hard (they walk away frustrated). Or auto-generation of terrain / features. Or automated testing where a RL agent basically plays the game a bazillion ways to find glitches. Or the generation of new animations from existing ones based on size / gender (no need to return to expensive mocap studios for additional data). That sort of thing.

[Elliecesa]

I'm astonished that I whatched a 27 minutes video and didn't understand a thing, I feel like I'm in school again

[onefulltimeequivalent]

22:18 This seems like a crappy argument to me. It was stated that the model is underperforming in practice because it was trained partly with non-available data. I wonder how you would ever even run the model without that particular data. I mean, when you assemble the input feature vector to feed through the model, what do you put in place of that now missing information? You would consciously run the model with missing information so there‘s no point really to look for a reason for its poor performance. Can someone please explain this to me?

[tacticslc]

Is there a patern to the video posted? Like art videos then programming... I've noticed that I think.

[theGreaterAwareness]

Sorry to be the critic but I wanted to hear from professionals that sound like they know what they are talking about.
I've finished every online course about NN that can be searched for and found. Almost all of these courses suffer from at least one of a few things.
1) Most of them are taught by a person that has no presentation skills.
2) Some of them are taught by people with presentation skills but do not understand NN.
The most common is #2 where we have people who just recite their own textbooks and try to sound like they know what they are talking about but they have ability to go in-depth for fear of talking about something they do not understand.
I really appreciate the time people put into these presentations but to not be critical is a disservice at the same time. I found this talk to be like an elaborated version of 2. This talk is for presentation purposes and some of the examples used abbreviations that were never explained - (requiring viewer thinking time which distracts from the talk).

The talk covered pretty much what every text book covers except without teaching anything - kind of like an overview of ML/NN. I do not know who this talk is focused on because it is too poorly presented to be suitable for absolute beginners. This presentation is too shallow to be useful for students and it is lacking anything useful such as "how does Havok deal with problems?" to be used by professionals.
I thought the female talker was the most interesting as finally there was something refreshingly different - "Changing the way you view your data so as to help find / locate possible reasons". But it was an overview only and I think it helps professionals but not anyone below because nothing was taught that a beginner can learn.
So I don't know what this talk is trying to achieve. Having said all this - it is a video from 2015 which explains a lot because much of the new students have surpassed the garbage in this video.

For people that do want to learn Machine Learning or Neural networks - I recommend two courses to start with.
Both of these are Udemy courses and you shouldn't be paying more than $15 for them so wait for them to be on sale and look for discount codes.
1) "The Visual Guide on How Neural Networks Learn from Data" - For absolute beginners - start here as this helps you visualize and know what is necessary and unnecessary.
2) "A Beginner's Guide To Machine Learning with Unity" - Once you have completed the first course, do this one if you have interest in Unity - which I think you should while you are studying as it allows you to focus on the code rather than all the other aspects required to make games/programs that can use all the latest video/audio/computer drivers.
After you finish these two courses, all the other courses will make sense to you with much less struggle from your brain.

[jasonzhou]

ML is kinda overkill and overwhelming for coding gaming AI in my opinion, but fantastic talk

[Semicolon]

I was hoping they would give examples of where you would use machine learning in games. Rather than a laundry list of techniques