Deep Learning

Just wondering if any one is actually using deep learning or neural nets in their game?

I’m planing on doing so, but based upon experience, it can be a bit of a rabbit hole (well it is lets face it!).

Does any on have any thoughts?

We used some kind of deep learning in one of our projects. And even though I am a big fan of machine learning and related AI concepts in general, in our case it turned out the efforts to implement and train was not worth it. Also games using such techniques for their AI are sometimes a bit hard to control in advance, i.e. the shipped game might not behave as you have intended it. That is not necessarily an issue, in particular if you build your game on that very aspect. But at least something to consider and keep in mind.

In our case btw. we didn’t implemented the deep learning system for UE4, it was for a custom engine …

This game is the only such thing I’ve seen lately in the community. Looks interesting.

I did some experimentation with neuroevolution a couple of years back and would love to try doing something with it in a UE4 project, but don’t have the time at the moment.

In my limited experience (license plate recognition), I had no idea whether it would work or not and it was very hit and miss and kind of ending up making up elaborate stories as to why it works or not. I’m pretty sure you can not debug a Deep Learning system to work out why it made a decision (despite what some presenters imply).

Just being thinking about

In some ways it seems more like you use Deep Learning because it is really fast (once it is trained) to come up with an AI decision and can save huge CPU cycles in making real time decisions and not really because it will actually save you any design work.

I kind of think you need that classical design (or some other system - humans maybe) to prove to others and yourself that the system works (which is fair enough). And you would also need yet another independent system to train.

An example of a classical system could be a Markov decision tree which would take much time to decide what to do and try to replace it with the Deep Learning system. If your classical system was simply based on simple heuristics, I’m not sure what the value of the Deep learning would be because it would be far simple to brute force the variables in your heuristics to find out what works the best.

The adapt to a player (as in the Kikai) sounds great and I have heard of the idea before, but I’m not too sure whether it would work in a convincing meaningful way.

To try to explain where I am coming from is that there is a classic case where you train a Neural Net to correctly perform an exclusive OR. How long it take to do this is very variable depending upon on your training variables (and data of course). So going back to a game case, this learning AI would most likely (murphy’s law) either take too long to learn (player gives up) or too short in which case it does not really appear that there is any learning going on. So I am thinking the risk is that perhaps 5% of the players would get this amazing experience of the the AI learning while 95% it is all a big yawn.

For me, maybe I should try to work out whether some one prefers to turtle or rush and kind of go from there perhaps.