Artificial Intelligence in Unreal Tournament 2004

[JACKAL]

readybot
http://robocup.rwth-aachen.de/readybot/

Strategic Planning for Unreal Tournament© Bots
http://www.aaai.org/Papers/Workshops/20 ... 04-005.pdf

A free book for download at MITPRESS
Artificial Life XII Proceedings of the Twelfth International Conference on the Synthesis and Simulation of Living Systems
http://mitpress.mit.edu/catalog/item/de ... &tid=12433




Artificial-Intelligence-in-Unreal-Tournament-2004
http://www.docstoc.com/docs/32527670/Ar ... ament-2004

LINK when Dead ........ will dig up something else
Creating .NET AI Bots for Unreal Tournament
http://channel9.msdn.com/coding4fun/art ... Tournament

[Ghosted]

Interesting stuff. Always thought it would be fun to try to improve bots by feeding demorecs/heat maps from good players into a machine learning engine. I'm sure Tim has done similar as I recall there was an article where he mentioned automatically adding bot pathing from heat maps.

The second link is completely out of date btw. Almost everything it links to is a 404 or a redirect to a default page. Probably better implemented in uscript anyhow.

[JACKAL]

Not sure why the link went dead, I just saw it yesterday befor I posted the link.
Its on You Tube also.
Microsoft sites never get deleated as there is no one to edit the database or the MS domain just to remove a few older sites.

Interesting stuff. Always thought it would be fun to try to improve bots by feeding demorecs/heat maps from good players into a machine learning engine. I'm sure Tim has done similar as I recall there was an article where he mentioned automatically adding bot pathing from heat maps.

The second link is completely out of date btw. Almost everything it links to is a 404 or a redirect to a default page. Probably better implemented in uscript anyhow.

[CTC-JimRimya]

I've always thought the bots on UT were pretty well programmed, save for a few things: They see you when you're invisible, and if you're the closest opponent, yet still on the other side of the map, they'll take shots at you with stupid weapons like the flak cannon or rocket. Although, sometimes I've been hit by a bot-fired rocket that locked onto me from across the map!

[Variasaber]

I've always thought the bots on UT were pretty well programmed, save for a few things: They see you when you're invisible, and if you're the closest opponent, yet still on the other side of the map, they'll take shots at you with stupid weapons like the flak cannon or rocket. Although, sometimes I've been hit by a bot-fired rocket that locked onto me from across the map!

There's this map I always liked, I think it was called Bania or something, basically Egyptian tileset or whatever, two sides, a huge chasm between them with water and a straight walkway at the bottom with elevators on either side of it to get up to the flag, some mantas, some other elevators that lead to secret rooms with powerups etc.
Anyway, if you play against bots and set them to a high difficulty, they spawncamp you FROM THEIR SPAWN with shock rifles. It's awful. And they'll shoot the mantas so that they slide into you and run you over, I dunno how they're smart enough to do that.

[JACKAL]

I think the cool thing about the bots is they have a learn mode when we play them. they nail us by the weak spots we have, or if we get the upper hand then they will "find" a way to get us anyways. this is cool stuff.

I messed with them alot over the years on a local server to see what gives.
if the bots have the time to "learn" gameside then they will work off each other too! I had a map with custom settings and the bots started to heal each other too. this was funny to see as one would turn and run off and heal a bot with low health. funny stuff. I had a few set to "think out loud" so I could hear what was going on gameside too! aslong as the bots are not set on 100% KILL, KILL, KILL. they will learn some cool moves. Not bad for AI on any game.

I've always thought the bots on UT were pretty well programmed, save for a few things: They see you when you're invisible, and if you're the closest opponent, yet still on the other side of the map, they'll take shots at you with stupid weapons like the flak cannon or rocket. Although, sometimes I've been hit by a bot-fired rocket that locked onto me from across the map!

There's this map I always liked, I think it was called Bania or something, basically Egyptian tileset or whatever, two sides, a huge chasm between them with water and a straight walkway at the bottom with elevators on either side of it to get up to the flag, some mantas, some other elevators that lead to secret rooms with powerups etc.
Anyway, if you play against bots and set them to a high difficulty, they spawncamp you FROM THEIR SPAWN with shock rifles. It's awful. And they'll shoot the mantas so that they slide into you and run you over, I dunno how they're smart enough to do that.

[Ghosted]

I think the cool thing about the bots is they have a learn mode when we play them. they nail us by the weak spots we have, or if we get the upper hand then they will "find" a way to get us anyways. this is cool stuff.

Do you have any documentation on this? Searching with queries like 'ut bots learn' pulled up nothing for the engine itself.

I did find a similar project for UT3:

http://ut3bots.codeplex.com/releases/view/18781

A few areas I'm interested in:

- Caching learned behaviors per map/player
- Offline learning
- Model building from player training data

[JACKAL]

I wish it was that easy of a search.
I found it (UT AI) at MIT site years ago. this is why i got on UT 2004.
I will dig up something, its hard because its all research papers at MIT or the like.
I always find this stuff searching Unreal 2004 AI or unreal 2004 bot AI, bluh bluh, bluh.

I think the cool thing about the bots is they have a learn mode when we play them. they nail us by the weak spots we have, or if we get the upper hand then they will "find" a way to get us anyways. this is cool stuff.

Do you have any documentation on this? Searching with queries like 'ut bots learn' pulled up nothing for the engine itself.

I did find a similar project for UT3:

http://ut3bots.codeplex.com/releases/view/18781

A few areas I'm interested in:

- Caching learned behaviors per map/player
- Offline learning
- Model building from player training data

[rimmer59]

I think the cool thing about the bots is they have a learn mode when we play them. they nail us by the weak spots we have, or if we get the upper hand then they will "find" a way to get us anyways. this is cool stuff.

If this is possible, which im sure it is with all of the new AI, it will not be as boring facing bots in instant action, or online either.

[Ghosted]

I think the cool thing about the bots is they have a learn mode when we play them. they nail us by the weak spots we have, or if we get the upper hand then they will "find" a way to get us anyways. this is cool stuff.

If this is possible, which im sure it is with all of the new AI, it will not be as boring facing bots in instant action, or online either.

It's possible, but there's an upper bound that's constrained by time. Maybe the bot can learn that you usually jump right and will therefore have a tendency to lead the shot that way. Learning that around corner x, you are likely to shoot 2 shock balls, with one as a feint is much less likely as it would require a more detailed model (which would increase time/complexity).

Again, interested on seeing what Epic did specifically around bot learning. Most of the code I've read is based on the bot navigating the environment in a convincing manner.

Here's an article on UT3 AI.

New features he mentions:
- weighted path traversal based on certain factors (ex: avoid killzones)
- creating a high level control interface, so a bot says 'move forward', which works similarly for most vehicles
- added more computationally intensive bot pathing with raycasting

None of these is really 'learning'. The first item could be described as 'follow the path of least resistance', where resistance along paths is changed by external factors. The bot generally follows the same instruction, it's 'brain' is not changing.

This book is a really great practical introduction to programming game AI and what it entails. It's very cool, but also very much 'nuts and bolts' stuff as Mark Rein alludes in the above article.

Realtime game AI that can really learn like you or I is still a bit away. Remember it's still gotta run at 60fps. They're still using supercomputers to simulate the neocortical columns of rats.

[JACKAL]

Yes, 100% with you.
You get my point of "learn? as for the bots.
Not that I had a lot to say, but its seeing what the bots are doing on the maps sometimes cracks me up, last night was good, this bot was coming after me and stoped and ran off to back track and circle back on me.
he took a few shots at me and zoomed off. he he
Two other bots "hid" till they had the health to fight me and then came and kicked my ass a few rounds. I was sticking it to them a few rounds just before.
I looked for them for maybe 5 mins, maybe more.

I will check the links out here soon too so I can maybe add a few things to my res. I could use it. hehe.
I was just thinking a few things over as for the settings on the UT 2004 server bots config. if the settings where to lean on the tactics and not the hit rate something should come of it then just a few fast kills. I may dig back in to this again as I am looking in to coding UT with Visual Studio 2010. I am loving it all over again.

Peace out Ghosted.
Thanks for the chit chat too!

I think the cool thing about the bots is they have a learn mode when we play them. they nail us by the weak spots we have, or if we get the upper hand then they will "find" a way to get us anyways. this is cool stuff.

If this is possible, which im sure it is with all of the new AI, it will not be as boring facing bots in instant action, or online either.

It's possible, but there's an upper bound that's constrained by time. Maybe the bot can learn that you usually jump right and will therefore have a tendency to lead the shot that way. Learning that around corner x, you are likely to shoot 2 shock balls, with one as a feint is much less likely as it would require a more detailed model (which would increase time/complexity).

Again, interested on seeing what Epic did specifically around bot learning. Most of the code I've read is based on the bot navigating the environment in a convincing manner.

Here's an article on UT3 AI.

New features he mentions:
- weighted path traversal based on certain factors (ex: avoid killzones)
- creating a high level control interface, so a bot says 'move forward', which works similarly for most vehicles
- added more computationally intensive bot pathing with raycasting

None of these is really 'learning'. The first item could be described as 'follow the path of least resistance', where resistance along paths is changed by external factors. The bot generally follows the same instruction, it's 'brain' is not changing.

This book is a really great practical introduction to programming game AI and what it entails. It's very cool, but also very much 'nuts and bolts' stuff as Mark Rein alludes in the above article.

Realtime game AI that can really learn like you or I is still a bit away. Remember it's still gotta run at 60fps. They're still using supercomputers to simulate the neocortical columns of rats.

[JACKAL]

I just found this bit of info.
http://robocup.rwth-aachen.de/old/ready ... /aiide.pdf
I did have some stuff saved from MIT.EDU site years go, but cant find it now.
MIT said at the time the AI in UT 2004 WAS the best the public could get in a game.
They got very complex very fast with the AI in UT 2004, blew me away. thats why I got UT 2004
I will keep on the MIT site for better info again

I think the cool thing about the bots is they have a learn mode when we play them. they nail us by the weak spots we have, or if we get the upper hand then they will "find" a way to get us anyways. this is cool stuff.

Do you have any documentation on this? Searching with queries like 'ut bots learn' pulled up nothing for the engine itself.

I did find a similar project for UT3:

http://ut3bots.codeplex.com/releases/view/18781

A few areas I'm interested in:

- Caching learned behaviors per map/player
- Offline learning
- Model building from player training data

[Ghosted]

There's a free VS plugin I use for UScript called nFringe

It adds syntax highlighting and importantly for me 'goto implementation'. I created a project and pointed to all of the UT3 source. Setup my build process to point to the ut compiler as a post build step and some other hacks to get it automated. I think there's better support through an actual project now, but the setup works and allows me to do scripts in VS and go peak at any of Epics uscript code at the touch of a button, which is invaluable.

Another thing I wished I'd learned sooner. There's a way to modify in game variables from the console. Forget the command, but saved a ton of time when I was trying to tweak values of vehicles, weapons, etc. It goes against the uscript object model, so you should be able to get at just about any object with it, including bots.

[JACKAL]

Yeah, I found nFringe, I have a ton of projects i need to use TFS
or some kind of source control,
i got to much stuff floating around my drives.
Thats what I am doing now, I am losing track of my work with out it.
The MSDN stuff I had worked well with demos of what the MS Dev did at
channel9 but it seems to have gone MIA on the MSDN site.

There's a free VS plugin I use for UScript called nFringe

It adds syntax highlighting and importantly for me 'goto implementation'. I created a project and pointed to all of the UT3 source. Setup my build process to point to the ut compiler as a post build step and some other hacks to get it automated. I think there's better support through an actual project now, but the setup works and allows me to do scripts in VS and go peak at any of Epics uscript code at the touch of a button, which is invaluable.

Another thing I wished I'd learned sooner. There's a way to modify in game variables from the console. Forget the command, but saved a ton of time when I was trying to tweak values of vehicles, weapons, etc. It goes against the uscript object model, so you should be able to get at just about any object with it, including bots.

[Helen]

I've always thought the bots on UT were pretty well programmed, save for a few things: They see you when you're invisible, and if you're the closest opponent, yet still on the other side of the map, they'll take shots at you with stupid weapons like the flak cannon or rocket. Although, sometimes I've been hit by a bot-fired rocket that locked onto me from across the map!

I always hated bots in raptors shooting from top sky of actioncity still not missing one shot...

Yeah bots have been programmed very well , but keep in mind bots do things what are according whats on the map. To truely utilize bots on a map really all on placement of (AI nodes defences, path nods, jumpspots, vehicle roads, snipe spots and exx..), but alot map makers dont really put the effort into those because map makers make maps for us to play mulitiplayer not single player and its ALOT of work.

There is one weakness to bots ui in ut2004 is "vehicles" a script would be nice not to have bots jump in and out every vehicle they get near .....