I see in my incoming blog links that someone linked my most recent post with the following comment:
Christer Ericson is upset about researchers pointing out that there has been little innovation in game AI since the 1980s. What do you think about this subject?
Ugh. As I dislike being misquoted, I guess I should spell it out by numbers in case anyone else doesn’t/didn’t get it:
- I’m not upset. If I got upset by academics saying stupid things I’d be the world’s angriest man (but I’m not).
- “Pointing out” means that there was something valid about the whine I quoted. As there wasn’t, the correct term here would be “claiming.”
- My previous blog post was about the incredible irony of someone whining about the state of game AI and how academic work is ignored, at the same time as their own academic work is, frankly, outright awful†. (But, hey, at least I didn’t ignore it, even though it deserved to be ignored.)
- I was also suggesting, based on personal experience, that the academic work that was implied being relevant to games (i.e. computational intelligence) is, contrariwise, nearly useless.
So, there, everything explained! Now back to the regular programming.
Addendum
†) Piss-poor AI (that wouldn’t even be suitable for an amateur game) aside, if you’re doing academic work on racing game AI, how can you possibly ignore testing your AI in existing frameworks like TORCS and RARS?
Well, thanks for the clarification!
alexjc
P.S. Nice headline again :-)
I blame the innertubes on reduction in reading skills. Who needs to read when you’ve got YouTube!
I was fishing around the internet for car-related AI and in particular feedforward multilayer-perceptrons when I came across Togelius’s blog and then jumped over to find what appeared to be a personal attack on academics. I think you have behaved just as badly in ‘misquoting’ by showing the worst of all 3 or so videos of the AI – the result which wouldn’t be so bad given that in context he was trying for more destruction derby type strategy.
Furthermore, you go to implicate the irrelevance of academic reserach in the industry when you should realise that Colin McRae, Forza and other car games (especially non-tarmac racing simulations) already use techniques such as MLP where heuristics/state based systems fail. Furthermore, companies with actual resources devoted to R&D, such as Valve and Crytek, do take the risk of developing and testing more complex and unscripted AI – something Togelius is trying to inspire.
You may have worked on a top-selling title that is based very much more on story-telling through art (I use that term loosely to involve script and game-specific code) instead of complex interaction between the player and the AI. That’s fine and I’m sure you’ll agree, making top-selling and fun games does not require implementing complex interaction with NPCs. But games like Nintedogs shows there is definately a gaping hole in the market.
Racing games push the bleeding edge in graphics, and now AI, because they are so easy to make and there has been so much competition between brands that they need someway to ‘move forward’ or to differentiate themselves in the market place. The same is now becoming true for the FPS genre and soon you could see advanced AI techniques spilling over to MMORPGs. There is only so far games can fake AI, but going to the next step requires companies with large resources to take risks. That is exactly what Togelius wants and the game-playing community has the right to demand it… after all, the customer is always right.
Jarrod
Jarrod, before working in games I was an academic (teaching AI, logic, etc) and doing research on genetic algorithms. Having had my feet roughly equally planted in both academia and in games for the last 20 years I can with reasonable accuracy identify what type of academic AI work is relevant to games and Togelius’ work just isn’t, nor is most everyone elses. There are very simple reasons for this irrelevance, including a near complete lack of understanding in academia of the process of making games and what our needs are, plus academia’s love of only ever solving toy problems, to name a few.
I don’t want to continue harping on the lack of quality or relevance of Togelius’ work because that was never the point (his ironic claims were), but to counter your point, note that, for games, the quality of an “AI” (or lack thereof) is defined by its worst behavior not its best behavior. As such, the clip I posted was highly relevant in showing just how poor Togelius’ results, and thus how ironic his hubristic statements, were.
Lastly, on the topic of irony… You imply some companies like Valve and Crytek are indeed hip to Togelius’ ideals, to cast aspersions on my claim of the irrelevance of most academic AI research to games. This is ironic because the person who posted in his reply on my first blog post that “I would not use as harsh words as Christer, but I agree with him” is an AI programmer at Crytek!
Okay,
Crytek don’t make racing games as far as I know and that Togelius used Multilayer Perceptrons (ie. MLP) in his work and there is already in fact a relevant and working implementation in Colin McRae Rally 2.0 which was made back in 2001. Just because Togelius hadn’t refined his work to release level didn’t say that it wasn’t refineable, Colin McRae has the final result:
http://www.generation5.org/content/2001/hannan.asp
Genetic algorithms were probably a bad choice for teaching his MLP (Colin used back-propagation, namely RPROP), but certainly he’s looking down the right path.
If you were to decide the AI for a rally game would you have thought that MLPs were pertinent to the task? Believe me, I hear what you are saying. At the end of the day a game needs to get past internal, publisher and console manufacture QA; the game mechanics have to work at all cost – and furthermore have to reach that Christmas sales season and be on budget. But the fact that you outright dimissed an approach to AI which is already actually proven-to-work in mainstream titles is to me somewhat of an error – no matter how much rhetoric is slapped on top.
If you are as discouraging as this in your lead role, perhaps AI ideas that might have one day been somewhat worthwhile would be quashed even before emerging?
But to your merit, indeed Togelius should have:
1) worked with a common and game-like framework like TORCS etc.
2) worked much more on teaching his AI to a production level before blowing his own trumpet
If more ‘games academics’ worked to this level of professionalism then perhaps they would have a greater contribution to the actual industry. But then, that’s closer to what you meant and what you should have constructively said in the first place…
Jarrod
“[…] But then, that’s closer to what you meant and what you should have constructively said in the first place.”
Jarrod, please don’t pretend to know what I meant or what I should or shouldn’t have said! It’s my decision to call things for what they are with or without providing constructive criticism. This is a blog, not a scientific paper; I expect my readers to be intelligent enough to appreciate the difference. I also expect them to be smart enough to provide and apply their own critical thinking to what is said, especially where I hint but deliberately omit detail.
I expect readers of this blog to understand that “because they sound cool” or “Colin McRae Rally 2.0 used it” are not valid justifications for using neural networks, and for them to be able to look up critiques of genetic algorithms or other computational intelligence paradigms (with bonus points given for reaching the conclusion that GAs are useless drivel, outperformed by simple stochastic hill-climbing schemes).
If this is too much to ask, then this blog might be a better destination.
Technique used in recent quality well-selling games (Forza Series, Colin McRae Series, Sprint Cars Series, I’m sure I can find more…) and will be used into the future = Relevant to the games industry.
“Because it sounds cool” was not the reasoning for using these techniques – there simply was no better way to do the job, especially in the rally/sprint cars titles. Just providing the turning, throttle/brake outputs for a car when they are often sliding around corners at a variety of differing entry speeds with also need to avoid colliding with the car in front… requires a complex formula. One that neural networks can determine or at least approximate using a number of hyper-planes.
Of course, we could just fix the cars to driving lines and have them fake the movements, but then when they collide and re-enter the track at different speeds, they will look stupid at best and of course in a reasonable simulation, the player wises up to the fact they are being taken for a ride.
Using an MLP the cars can drive by themselves and with decent training can cope with new tracks without some QA guy retweaking coefficients or needing to fiddle with track data so that cars behave as they should. It’s a saving in time on the art-side and its also an extra layer of realism that seems to be now industry standard.
The Forza guys took it a step further to add a fancy dot-point to their box, Drivatars. Using bayesian learning (this is a guess, given the AI programmer’s CV I found on the net) it lets the player driver around the track for a while, analyses his/her driving style and makes a new AI driver that can allow the player to leave it go for a while. So if the player is more interested in the management side of racing, they can focus on that…. though in the case of Forza, its just a dot-point where the potential customer goes ‘cool’ and buys it instead of the other product without the dot-point.
In defense of Togelius, I actually decided to revisit his page and *gasp* look at his papers. I found that, despite the ill-conceived competitive AI which you are so fond of, his ‘regular’ AI can play ‘better than human’ and be generalised to deal with a wide variety of tracks (but a couple of really nasty ones). Also, genetic algorithms aren’t necessarily worse than gradient descent methods of backpropagation, just a hell of a lot slower. But they could be less likely to get stuck on local minima on the error surface given a good training pattern.
But returning to your original article you did write of computational intelligence as a big wet dream, and I do agree that it isn’t a magic bullet that makes games better (Black and White would be a thoroughly good example of this). But at the very least, for racing games, neural network techiniques have been used in a number of top-selling car racing simulation to increase realism.
Such innovation is scorned upon less in graphical realms, such as perhaps papers on spherical harmonics (I recalled SCEA wrote one?) inspiring new techniques in real-time lighting. I guess because innovation in AI *is risky* as Togelius pointed out, the industry is so averse to trying new things in NPC interaction. AI behaviour directly affects gameplay and if we make agents too intelligent the QA process becomes broader and less defined… and to what benefit to the publisher in sales? Huge risk, indeed, but there’s also an approaching risk of gamers getting bored with the same old crap in every game.
Anyway, I hope you had a good Christmas between my tedious responses, and should I not get the chance to say, all the best for the new year.
Cheers,
Jarrod
I know there are commercial racing games using neural networks. You’re not telling me anything new, honestly. Your statement that there’s “no better way to do the job” reflects more on you than anything else. To suggest that anything other than neural networks would imply cars going on rails is either ignorant, dishonest, or both (pick whichever suits you). There is absolutely nothing “innovative” about using neural networks; they’ve been around for ages (and you can find machine learning approaches used in BASIC game listings from the 1970s). As a generic statement, they (NNs), like other computational intelligence paradigms, are almost always the wrong tool for what we do.
I don’t know why you keep harping about Togelius’ work. If it’s not clear already: he has shown nothing relevant to real games as he worked with a toy problem that uses about none of the parameters of a real game. You can’t show results about oranges by working with lemons.
“SCEA” is an organization. Organizations don’t write papers, individuals do.
Happy holidays to you too! I’m looking forward to the next long post that your misreading of and jumping to conclusions about something that I just wrote (or excluded from writing) will trigger! Though there is something to be said for both preciseness and conciseness.
Hehehe, yeah, this thread could go on forever and I’ll certainly agree its diverged far from the track where everything started. My original gripe is that you really do launch into quite personal attacks and name-calling rather than laying down a logical argument – though now you really have fleshed out your argument, but it took quite some coaxing.
Okay, as it stands, so you agree that neural networks in car games are a sensible approach. While they are most suited to dirt-track simulation, I know they aren’t the only approach. I certainly know they aren’t ‘innovative’ which is something I was trying to point out to you by showing you some 7+ years of usage history. The reason I was pointing this out to you is because you mentioned something vague about ‘knowing what works’ in your 20 years of experience as to why Togelius’ system is doomed to fail, but he also uses MLP and had made reasonably robust neural network. He also made a not so robust experiment which you turned into a tool of ridicule.
Going before that, to the original blog post you imply that Togelius chose to use this method of ‘computational intelligence’ of MLP was purely academic and ‘toy’ (and have futher stated in this thread). Yet games in the real world do use this technique, something which you only ceded to in the last reply like you had it in mind all along.
But you’re right, I’m still straying for the point here, what I found profoundly disturbing is that given your influential voice, and a very articulate one at that, rather than providing a constructive critique you said to your buddies ‘come laugh at this fellow’ and denegrated to name-calling and such. He did unwittingly do you some insult, but a more informative approach and less bullying would be something I’d expect more from a veteran of the industry.
I felt the need to stick up for the guy and I did. By the way, do enjoy reading the rest of your blog. Hope you enjoyed your holidays.
Cheers,
Jarrod
Jarrod, I honestly cannot take more of your dimwitted posts here. You half-read what I’ve written, completely reinterpret it in your own mind, and then form illconceived arguments in your posts about the resulting strawmen. Nevertheless I’ve tried to be reasonable and reply to your posts, but it’s clear to me that I’m just wasting my time replying, as you cannot follow a simple syllogism to save your life (as evidenced by just about every line in your last post).
I would greatly appreciate if you could take your childlike energy and thinking elsewhere. I assure you, this is not the blog for you.