One thing to keep in mind is that Gladiator/Q3A bots followed very simple patterns and never played like a human would. Over the time, modders made their own bots that were much more human-like.
Spiterbot was a notorious one. It was capable of strafejumping, rocketjumping, zone denial, pickup timing, risk/reward behavior, enemy power estimates, and all things you would expect from a competent human player. Things like pro-q3dm6 bridge-to-railgun jump weren't a problem for it. It was still too easy to read for a good player, and susceptible to mistakes leading to HP/armor starvation, like avoiding the railgun control line like plague where a human would have tried to outsmart an enemy, risked it, or waited for an enemy mistake. Too bad there never was a technical write-up for the Spiterbot like this one.
(also, the article is from 2001, this should probably be in the title)
To elaborate slightly, it wasn’t just a written report; it was his Master’s thesis. Therefore, there was undoubtedly a significant incentive to complete it.
- navigation, which is done with a combination of A* or Dijkstra and navmeshes
- decision, which is done with the paradigm of behavior trees
- computer vision, to complete what navmeshes can not provide
Despite the term "computer vision", it is still not AI.
I will quote the "Navigation" and "Computer vision" section from the site I linked above:
> All bots should be able to reach any part of the map that does not require special training.
> Bots should be able to detect any enemy in their FoV. It would be nice to have bots not being too good at finding hidden buildables on the roof. This makes hiding eggs unfun and next to impossible in the current state. Unless they are placed outside of a navmesh, in which case bots are totally unable to discover them.
Must be semantics because I don't consider bots within this context AI. Just like I don't consider web crawlers AI. Just like I don't consider an IRC bot AI. Just like I don't consider Quake 3 bots AI.
I mean... do you consider an IRC bot AI? "It is a bot, therefore it is AI" seems wrong to me.
You're making my point (game AI != ML-based AI). If "Game AI" needs to be distinguished from ML/actual AI, then it's not AI, it's just what we've historically called rule-based logic.
The original claim was "Bots are AI, always have been". I gave counterexamples: IRC bots, web crawlers, Quake bots. None of these involve intelligence, they're deterministic programs following explicit rules.
Your "game physics" analogy is also weak. Physics simulation approximates real physical laws. Calling it "physics" makes sense, it's modeling physics. But pathfinding algorithms and decision trees aren't modeling intelligence; they're just conditional logic. That doesn't make it accurate or correct, just conventional.
We need to stop calling things that aren't AI, AI, especially today.
One thing to keep in mind is that Gladiator/Q3A bots followed very simple patterns and never played like a human would. Over the time, modders made their own bots that were much more human-like.
Spiterbot was a notorious one. It was capable of strafejumping, rocketjumping, zone denial, pickup timing, risk/reward behavior, enemy power estimates, and all things you would expect from a competent human player. Things like pro-q3dm6 bridge-to-railgun jump weren't a problem for it. It was still too easy to read for a good player, and susceptible to mistakes leading to HP/armor starvation, like avoiding the railgun control line like plague where a human would have tried to outsmart an enemy, risked it, or waited for an enemy mistake. Too bad there never was a technical write-up for the Spiterbot like this one.
(also, the article is from 2001, this should probably be in the title)
To elaborate slightly, it wasn’t just a written report; it was his Master’s thesis. Therefore, there was undoubtedly a significant incentive to complete it.
J.M.P. Van Waveren was Carmack's right hand man for many years, passed way too soon.
https://mrelusive.com/
We had bots in Tremulous which is an old ioquake3 fork, there was no AI involved, worked fine. See https://github.com/cmf028/CoW-Tremulous.
Unvanquished (https://unvanquished.net) became the new "Tremulous" (different engine, etc.), also has bots.
Read more about it here: https://wiki.unvanquished.net/wiki/Bot_design
- navigation, which is done with a combination of A* or Dijkstra and navmeshes
- decision, which is done with the paradigm of behavior trees
- computer vision, to complete what navmeshes can not provide
Despite the term "computer vision", it is still not AI.
I will quote the "Navigation" and "Computer vision" section from the site I linked above:
> All bots should be able to reach any part of the map that does not require special training.
> Bots should be able to detect any enemy in their FoV. It would be nice to have bots not being too good at finding hidden buildables on the roof. This makes hiding eggs unfun and next to impossible in the current state. Unless they are placed outside of a navmesh, in which case bots are totally unable to discover them.
There is a pull request that implements the use of "beacons" for bots which is used to locate buildables: https://github.com/Unvanquished/Unvanquished/pull/2683
There was no "computer vision" in the Tremulous bots, FWIW.
In Tremulous, they used waypoints only, so has nothing to do with AI either.
Bots in Unvanquished are quite human-like and there are plans to add personalities (if they did not add them already) and there is no AI there either.
TL;DR: You can have bots that work quite damn well and human-like without AI/ML.
Semantics? Bots are AI, always have been.
Must be semantics because I don't consider bots within this context AI. Just like I don't consider web crawlers AI. Just like I don't consider an IRC bot AI. Just like I don't consider Quake 3 bots AI.
I mean... do you consider an IRC bot AI? "It is a bot, therefore it is AI" seems wrong to me.
Is game physics physics? Game AI is a well established term, and it's weird to see it being confused with anything ML-related.
You're making my point (game AI != ML-based AI). If "Game AI" needs to be distinguished from ML/actual AI, then it's not AI, it's just what we've historically called rule-based logic.
The original claim was "Bots are AI, always have been". I gave counterexamples: IRC bots, web crawlers, Quake bots. None of these involve intelligence, they're deterministic programs following explicit rules.
Your "game physics" analogy is also weak. Physics simulation approximates real physical laws. Calling it "physics" makes sense, it's modeling physics. But pathfinding algorithms and decision trees aren't modeling intelligence; they're just conditional logic. That doesn't make it accurate or correct, just conventional.
We need to stop calling things that aren't AI, AI, especially today.
Awful paper. Almost zero citations!
> The ultimate goal of AI is to create an artificial man
Close tab.
To expand on this: https://news.ycombinator.com/item?id=45924721
No AI required whatsoever!