Artificial Intelligence Depot
Visiting guest. Why not sign in?
News, knowledge and discussion for the AI enthusiast.
FEATURES COMMUNITY KNOWLEDGE SEARCH  
Robot Building for Beginners
Not only does this book assist you in understanding component parts of robot development, but also it prepares you with techniques to learn new discoveries on your own.
More information at Amazon US UK

Reply to Message

Not registered yet?

The AI Depot has a focused community of friendly users. Rather than let anyone abuse the site at the brink of promiscuity, we prefer to let only those with an active interest participate... this simply requires registering.

Why not sign up!

Joining the site's community is completely free. You can then post messages freely, and customise your personal profile at will. Specific privileges will also be granted to you, like being able to access printer-friendly articles without restrictions. So, why not register?

Username:
Password:
Subject:
Email me when someone replies.
Body:

Parent Message

GDC AI Roundtable

I noticed that Ferretman has posted a review of the AI round-table at the Game Developers Conference over on his Game AI website; AI Roundtable Moderator’s Report. The page is very interesting to read and well summarised.

""Learning generated a lot of discussion, if not a lot of optimism [...] but nobody had a clear idea on how to keep the AI from learning something dumb.""

I have to admit quite a few points chilled my bones a bit, including this attitude towards learning. The fright about getting dumb AI is unfounded, as you can set objective functions to evaluate the performance of the artificial player -- guiding the learning, or restoring old versions. Besides, I think there's something wrong when the many benefits of learning are dismissed due to their so-called "unpredictability", and their impact on the testing procedure. More on that soon hopefully.

935 posts.
Sunday 14 April, 16:37
Reply
When will games learn?

I too am disappoint with this attitude towards on-line learning in games. There are very simple way to ensure that learning is effective in games, the most simple being to test for effectiveness during the game itself.

I am a little confused why learning is frowned upon and then other technologies that produce emergent behaviours, i.e. flocking and steering behaviours, are promoted. Learning seems to me to be treated unfairly in this regard.

I am also a little disappointed that learning is considered synonymous with neural networks and genetic algorithms. Powerful as these technologies may be they are not the only ways to implement learning, e.g. Q-learning, classifier systems, decision trees to name but a few.

It seems that there is a certain amount of ignorance behind the unwillingness of games developers to consider more advanced AI. I'd be interested to know whether the same developers are happy at the thought of building the very large and very complex FSMs that will be needed in the games 5 years from now.

Lets hope that the machine learning community and the games developers can conduct a useful dialog soon so that appropriate learning technologies can be identified.

Sorry... felt like a bit of a rant. :-/

-RoB-

47 posts.
Sunday 14 April, 18:00
Reply
re

yup, I was pretty horrified too. I was discussing with Steve how developers attitudes towards AI reminds me of the attitude UK investors have towards people like Dyson (of vac fame)and other innovative thinkers/inventors. They just haven't the balls to take the risks which may ultimately lead to a huge payoff.

I think largely, ignorance is to blame. Developers just don't seem to know what's going on in the AI scene and what current technology is capable of.

Our day will come soon though chaps...

51 posts.
Monday 15 April, 20:03
Reply
for learning AI to be appropriate in game development...

"Developers just don't seem to know what's going on in the AI scene and what current technology is capable of."

Perhaps. But they often have a number of good reasons to forego learning AI that haven't been mentioned here.
RoB put it nicely "[To] ensure that learning is effective in games...".

For learning AI to be effective in the overall creation of a commercial game (let's assume a console game, because that's the biggest market), the following should hold:
- the learning solution shouldn't be more expensive (resource wise) than the manually developed one; however, learning involves overhead
- the level designer should remain in control, because it typically is the level designer that directs most of the game play; learning AI takes control away from him (and makes it harder for him to design the level)
- developing the learning solution should not take much more time than writing some standard behavior with a couple of variants (few players notice the difference)
- the learning AI should not consume a larger bite of the test budget (which is tough to achieve, because you now need to test against AI that grows undesired behavior)
- the learning AI should not be more difficult to debug (and understand);
- the learning AI should not be more difficult to develop (and here you might blame the AI developers for not being up to date with the latest techniques).

If a 20 men development team is relying on you to come up with solid AI, on schedule, within the budgets (memory, ms / frame, developer effort, test effort), investing in a learning AI (vs a hard-wired traditional AI) isn't an obvious thing to do, unless part of the game design explicitly calls for it.
Even then, the learning part is a risk: will the results be good enough, or should we prepare for a manually designed fall-back?

In other words: yes, probably there are a number of game developers who are not willing to consider learning AI. But, on the other hand, for learning AI to be appropriate, it needs to address more game development concerns than just the AI theory.

William

19 posts.
Tuesday 16 April, 05:09
Reply
re

Hi William,

Although I agree with most of what you say I will question several points.

"- developing the learning solution should not take much more time than writing some standard behavior with a couple of variants (few players notice the difference)
- the learning AI should not consume a larger bite of the test budget (which is tough to achieve, because you now need to test against AI that grows undesired behavior)
- the learning AI should not be more difficult to debug (and understand);"

why not? If the AI makes the game a much better game it will make(if marketed correctly) more money. This is the bottom line after all. Creatures didn't sell a million copies because they were cute, it sold a million copies because people are able to interact in a meaningful way with the AI(and because they were cute ;0)). Same goes for the Sims... More revenue means you can afford to spend more time in development. But, I accept, there is a big risk to be taken initially.

The points you make seem to me to be more valid for traditional game formats and for working within a traditional development team. I see a future with new gaming experiences on the horizon.

ps. Great paper you did on terrain reasoning btw

51 posts.
Tuesday 16 April, 07:06
Reply
why not :)

There are two kinds of the "learning AI makes a better game".

The first kind is: the game design defines that the AI should learn. Think Creatures, Black & White, the Sims, a small part of the new Virtua Fighter 4.

The second kind is: the AI designer feels that by adding some 'learning' AI somewhere, he might achieve a more interesting AI. Sometimes this contradicts with the game design (the Kurt Warner or Ruud van Nistelrooij character should not adapt its way of playing football/soccer). In almost all cases, it should be balanced against the wishlists of the other team members (engine designers, level designers, etc.) whose ideas might have a larger impact on the overall game play.

Currently, I'm involved in an unannounced console game that features a large dose of AI. Maybe a tad too much AI to realize before the first milestone. On my planning, there are dozens of standard features that need to be realized in a "good enough" fashion. A few features have been earmarked as 'selling points' and require more attention. Then, there are features that require working together (and bargaining) with level designers and engine developers to get it right. Finally, there's my wish list (and the wish list of the full-time other AI developer) of things we'd like to add as well.
Both these wishlists happen to include some learning stuff. In my case, there's some reinforcement learning I'd love to see in the game. But I expect it will be difficult to convince the team/producer to add it, since it increases the testing effort. Perhaps I can find the time to just implement it, just show it, and see what happens.
The other learning item involves interpreting a complex situation. We don't have a good algorithm for that, and contemplate to use some GA/NN and player/tester feedback. Hopefully, that leads to more insight (and ideally a hard-wired algorithm for in the game).

I'm definitely not representive for other game AI developers, but perhaps this helps to understand why not every game includes learning AI. (Although more AI's were constructed with learning algorithms during production - for example, QuakeIII used GA's to tune the different AI personalities).

William

19 posts.
Tuesday 16 April, 09:55
Reply
Games and Concepts

When you play against other human players, they do things that you may not like, things that the designer didn't plan, things that look stupid, things that can break the engine code. But people put up with it, since as a whole, they can maximise their experience by avoiding the dodgy characters, and playing with their friends.

Integrating proper AI into a game should take that policy, a more relaxed attitude. Not all AI characters will suit the designer, though many things can be done to maximise the feel he's trying to create. Gaming experience will emerge if appropriate learning criteria is given to unique embedded creatures.

Likewise, quality of AI is a consequence of having unique individuals in the world. Some may be as dumb as your dumbest player, but if you design it well, chances are some will learn to do really original, unanticipated things... this provides amazing depth to the game, and as your environment becomes more interesting, so will the AI. Black & White is a simple environment, yet it has a few hours of gaming in it (and a few interesting moments). Picture that, but hundred fold.

I just think the entire attitude is wrong. Debug and test the engine, not the game. The experience is a consequence of putting everything together.

935 posts.
Wednesday 17 April, 09:07
Reply
re

I couldn't agree more.

51 posts.
Wednesday 17 April, 09:57
Reply

Back to the Artificial Intelligence Depot.