Neural Nets in Games
Author
Does anybody know the current state of play for embedding neural networks within games? Has anybody played around with this recently?
Mark Allen
Lead Programmer
Tashco Studios
http://www.tashco.com
I''d advise you to look at the Game AI page .
As for my personal experience, I am currently working on an game AI toolkit and the only uses we have found for NN was to model trust between NPCs or imitating evaluation functions for strategy planning.
Be reading you,
David
As for my personal experience, I am currently working on an game AI toolkit and the only uses we have found for NN was to model trust between NPCs or imitating evaluation functions for strategy planning.
Be reading you,
David
Without actually having played around with it I will just make a general comment on NNs.
A lot of people has over the years and have found them very hard to use in solving general purpose problems - including games.
The main problems is that NNs have to be trained and preferable by the the developer because runtime training by very well go bad and cause the NN to act stupidly. Furthermore, training and using and NN is quite CPU consuming compared to rule based AIs (including state machines).
Assuming that we are unable to simulate the number of neuron in the human brain, NNs are mainly good at solving problems that is somewhat simple in nature while being hard to define using rules. Campared to hard coded AI routines thay are often not so efficient.
The same arguments generally goes for genetic algorithms.
At a GDC ''99 roundtable about Game AI NNs were discussed and all current developers (as far as I remember) did not use NNs because it was to unwieldable.
Beyond this point there is a problem with NNs: They are not very reliable. This means that they usually act like a black box which makes the hard to test and debug before a release. So my suggestion to you is: Do not use NNs for games!
B.Sc. Jacob Marner
Graduate Student of Computer Science, The University of Copenhagen, Denmark.
http://fp.image.dk/fpelisjac/rolemaker/
A lot of people has over the years and have found them very hard to use in solving general purpose problems - including games.
The main problems is that NNs have to be trained and preferable by the the developer because runtime training by very well go bad and cause the NN to act stupidly. Furthermore, training and using and NN is quite CPU consuming compared to rule based AIs (including state machines).
Assuming that we are unable to simulate the number of neuron in the human brain, NNs are mainly good at solving problems that is somewhat simple in nature while being hard to define using rules. Campared to hard coded AI routines thay are often not so efficient.
The same arguments generally goes for genetic algorithms.
At a GDC ''99 roundtable about Game AI NNs were discussed and all current developers (as far as I remember) did not use NNs because it was to unwieldable.
Beyond this point there is a problem with NNs: They are not very reliable. This means that they usually act like a black box which makes the hard to test and debug before a release. So my suggestion to you is: Do not use NNs for games!
B.Sc. Jacob Marner
Graduate Student of Computer Science, The University of Copenhagen, Denmark.
http://fp.image.dk/fpelisjac/rolemaker/
I work with Neural Nets in a real world environment, we use them for fraud detection on telephone networks.
http://www.fraud-solutions.com/fraudsolutions/index.html
I''ve recently been considering NNs for game development and I think there could be uses for them but I''m not sure how successful they would be. The application I''ve been considering is NPCs in an online RPG which have the ability to learn. To implement an NN successfully you need to narrow down the problem domain, there''s no point in trying to have an NPC NN which controls every aspect of the NPC''s interaction with the world around it. E.g. what you could do is implement the NN so it performs the decision making for combat, you would need to train the NN up to a certain point but allow the NN to keep learning after the basic training, then the more battles the NPC takes part in the better the NN will respond to actions in battle.
I think I can get this to work, what I doubt I can do is make it feasible due to processing power, you would not be able to process more than a few NPCs on a single PC.
Cheers
Nigel
http://www.fraud-solutions.com/fraudsolutions/index.html
I''ve recently been considering NNs for game development and I think there could be uses for them but I''m not sure how successful they would be. The application I''ve been considering is NPCs in an online RPG which have the ability to learn. To implement an NN successfully you need to narrow down the problem domain, there''s no point in trying to have an NPC NN which controls every aspect of the NPC''s interaction with the world around it. E.g. what you could do is implement the NN so it performs the decision making for combat, you would need to train the NN up to a certain point but allow the NN to keep learning after the basic training, then the more battles the NPC takes part in the better the NN will respond to actions in battle.
I think I can get this to work, what I doubt I can do is make it feasible due to processing power, you would not be able to process more than a few NPCs on a single PC.
Cheers
Nigel
I have reason to make a NN that learns... processing power really isn''t too big of a barrier. I just need to know what to do with the dang neurons once I code them! How do I make it learn something other than simple math?
--------------------
You are not a real programmer until you end all your sentences with semicolons;
--------------------
You are not a real programmer until you end all your sentences with semicolons;
Visit the ROAD Programming Website
One of the major problems of using ANNs for anything is figuring out how to present the data to the system in the most usable form.
Should you feed every character and his position into the net? Should you use a 2D input grid that is fed threat values, or a threat/xpos/ypos neuron triplet for each enemy?
I think that once an effective form is discovered, ANNs do their job efficiently and effectively.
Should you feed every character and his position into the net? Should you use a 2D input grid that is fed threat values, or a threat/xpos/ypos neuron triplet for each enemy?
I think that once an effective form is discovered, ANNs do their job efficiently and effectively.
Hi,
I am working on NN, Its for a game called TNW. ( check it out at http://www.fortunecity.com/skyscraper/application/1003
and http://tnw.ezain.com
Neuronet is a good thing to add in a game. But to much of it can slow down the game dramatically. The number of calculations will be to much for the computer to handle and it can slow down internet gaming. But some NN in a game can be very usefull. Creatures will be more realistic and would react like real. NN is the future, so keep working on it
cya
I am working on NN, Its for a game called TNW. ( check it out at http://www.fortunecity.com/skyscraper/application/1003
and http://tnw.ezain.com
Neuronet is a good thing to add in a game. But to much of it can slow down the game dramatically. The number of calculations will be to much for the computer to handle and it can slow down internet gaming. But some NN in a game can be very usefull. Creatures will be more realistic and would react like real. NN is the future, so keep working on it
cya
This topic is closed to new replies.
Advertisement
Popular Topics
Advertisement
