GalCiv II: Computer AI debugging

Back a couple of years ago there was a great deal of discussion on Usenet
with regards to the GalCiv I AI. Most people thought it was pretty good but
others felt the military tactics were bad. That while the AI provided a
challenge, it mostly was able to do so because it was more effective at
managing its economy/industry than the typical human and thus could
essentially "win" through brute force.

In general, my view was that the AI worked pretty well but militarily was
disappointed. The main issue was its lack of coordination in projecting
military power into a given quadrant. The result was hat looked like the
age-old "Death train" (That's where units uselessly throw themselves at a
wall dying with no purpose). There simply wasn't enough analysis on putting
units together and using them effectivley.

A lot of this came down to being restricted to CPU light algorithms. GalCiv
I was a sprite-based game -- no hardware acceleration. So the graphics were
CPU bound. Even though the AI is, and continues to be, multithreaded, there
are still very real limits on how much time can be spent generating
multi-turn strategies. GalCiv II, however, is 3D-accelerated which has
freed up the CPU a great deal. This has allowed for greatly expanded
algorithms during military endeavors without raising the hardware
requirements (in fact, on the typical PC, GalCiv II will run better than
GalCiv I did ironically despite immense graphics improvements and 3D
engine).

A lot -- A LOT -- of work has been going into the AI of Galactic
Civilizations II in order to not just make it "challenging" but to make it
play intelligently. There were significant chunks of the GalCiv I AI that
were scripted. That is, the AI is programmed to do certain things due to
play-testing over and over. This time around, we're trying our best to make
the AI build strategies based on analysis and not rely on scripts.

The improvements in the AI have led to a lot of changes, probably the
biggest one being the elimination of the infamous "The AI knows where the
good planets are" issue that GalCiv from the OS/2 days on up to the 2003
release of GalCiv. GalCiv II's AI has to build scouts and look for good
planets just like everyone else.

To get an idea of what is involved in debugging such an AI, I chronicled one
such game. Mind you, we're now doing several of these types of games each
day to tune the AI, to find flaws, etc. The idea being that when the game's
released, the difficulty levels will determine how smartly the AI plays
rather than age old "Here's some money (or less of)" way that strategy games
have used as a crutch.

Anyway, here's the URL to the AI debugging session entitled "Human/Drengin
wars". I through in flavor text as well to make it more interesting to
read.

http://www.galciv2.com/Journals.aspx?AID=98513

Brad
--
Brad Wardell
Project Manager: Galactic Civilizations II (www.galciv2.com)
Stardock - http://www.stardock.com

"Brad Wardell" <bwardell@stardock.com> wrote in
news:yZudnZa7eZRE8UneRVn-og@comcast.com:

>
> A lot of this came down to being restricted to CPU light algorithms.
> GalCiv I was a sprite-based game -- no hardware acceleration. So the
> graphics were CPU bound. Even though the AI is, and continues to be,
> multithreaded, there are still very real limits on how much time can be
> spent generating multi-turn strategies. GalCiv II, however, is
> 3D-accelerated which has freed up the CPU a great deal. This has
> allowed for greatly expanded algorithms during military endeavors
> without raising the hardware requirements (in fact, on the typical PC,
> GalCiv II will run better than GalCiv I did ironically despite immense
> graphics improvements and 3D engine).
>
> A lot -- A LOT -- of work has been going into the AI of Galactic
> Civilizations II in order to not just make it "challenging" but to make
> it play intelligently. There were significant chunks of the GalCiv I
> AI that were scripted. That is, the AI is programmed to do certain
> things due to play-testing over and over. This time around, we're
> trying our best to make the AI build strategies based on analysis and
> not rely on scripts.
Nice update! I'm looking toward the game even more.

I assume that AI of each race can run independently so it should be
relatively easy to offload these calculations on other computers. Is there
a chance of this happening? - That would take care of lack of CPU power
nicely... In this case difficulty could be called "Intelligent, and yes,
you can use computing grid"

Alex.

"Brad Wardell" <bwardell@stardock.com> wrote in
news:yZudnZa7eZRE8UneRVn-og@comcast.com:

>
> A lot of this came down to being restricted to CPU light algorithms.
> GalCiv I was a sprite-based game -- no hardware acceleration. So the
> graphics were CPU bound. Even though the AI is, and continues to be,
> multithreaded, there are still very real limits on how much time can be
> spent generating multi-turn strategies. GalCiv II, however, is
> 3D-accelerated which has freed up the CPU a great deal. This has
> allowed for greatly expanded algorithms during military endeavors
> without raising the hardware requirements (in fact, on the typical PC,
> GalCiv II will run better than GalCiv I did ironically despite immense
> graphics improvements and 3D engine).
>
> A lot -- A LOT -- of work has been going into the AI of Galactic
> Civilizations II in order to not just make it "challenging" but to make
> it play intelligently. There were significant chunks of the GalCiv I
> AI that were scripted. That is, the AI is programmed to do certain
> things due to play-testing over and over. This time around, we're
> trying our best to make the AI build strategies based on analysis and
> not rely on scripts.
Nice update! I'm looking toward the game even more.

I assume that AI of each race can run independently so it should be
relatively easy to offload these calculations on other computers. Is there
a chance of this happening? - That would take care of lack of CPU power
nicely... In this case difficulty could be called "Intelligent, and yes,
you can use computing grid"

Alex.

"Brad Wardell" <bwardell@stardock.com> wrote in
news:yZudnZa7eZRE8UneRVn-og@comcast.com:

>
> A lot of this came down to being restricted to CPU light algorithms.
> GalCiv I was a sprite-based game -- no hardware acceleration. So the
> graphics were CPU bound. Even though the AI is, and continues to be,
> multithreaded, there are still very real limits on how much time can be
> spent generating multi-turn strategies. GalCiv II, however, is
> 3D-accelerated which has freed up the CPU a great deal. This has
> allowed for greatly expanded algorithms during military endeavors
> without raising the hardware requirements (in fact, on the typical PC,
> GalCiv II will run better than GalCiv I did ironically despite immense
> graphics improvements and 3D engine).
>
> A lot -- A LOT -- of work has been going into the AI of Galactic
> Civilizations II in order to not just make it "challenging" but to make
> it play intelligently. There were significant chunks of the GalCiv I
> AI that were scripted. That is, the AI is programmed to do certain
> things due to play-testing over and over. This time around, we're
> trying our best to make the AI build strategies based on analysis and
> not rely on scripts.
Nice update! I'm looking toward the game even more.

I assume that AI of each race can run independently so it should be
relatively easy to offload these calculations on other computers. Is there
a chance of this happening? - That would take care of lack of CPU power
nicely... In this case difficulty could be called "Intelligent, and yes,
you can use computing grid"

Alex.

In article <yZudnZa7eZRE8UneRVn-og@comcast.com>, bwardell@stardock.com
says...
> In general, my view was that the AI worked pretty well but militarily was
> disappointed. The main issue was its lack of coordination in projecting
> military power into a given quadrant. The result was hat looked like the
> age-old "Death train" (That's where units uselessly throw themselves at a
> wall dying with no purpose). There simply wasn't enough analysis on putting
> units together and using them effectivley.

An invariable problem, I suspect. You need strongish AI to strike a
balance between opportunistic raiding/exploration/dispersed defence and
attacks in force.

Part of the solution is often to change the rules. Civ4 did away with
separate attack and defence parameters, and instantly the cost to the
AI of ill-judged cavalry raids was reduced, though not eliminated.

> A lot -- A LOT -- of work has been going into the AI of Galactic
> Civilizations II in order to not just make it "challenging" but to make it
> play intelligently. There were significant chunks of the GalCiv I AI that
> were scripted. That is, the AI is programmed to do certain things due to
> play-testing over and over. This time around, we're trying our best to make
> the AI build strategies based on analysis and not rely on scripts.
>
> The improvements in the AI have led to a lot of changes, probably the
> biggest one being the elimination of the infamous "The AI knows where the
> good planets are" issue that GalCiv from the OS/2 days on up to the 2003
> release of GalCiv. GalCiv II's AI has to build scouts and look for good
> planets just like everyone else.

That was more of an issue because some people were claiming that the AI
played by exactly the same rules, I suspect ;-)

Of course, some people just don't understand the difficulty of writing
an AI and will always complain when sleight of hand is used to create
an effective computer opponent.

> To get an idea of what is involved in debugging such an AI, I chronicled one
> such game. Mind you, we're now doing several of these types of games each
> day to tune the AI, to find flaws, etc. The idea being that when the game's
> released, the difficulty levels will determine how smartly the AI plays
> rather than age old "Here's some money (or less of)" way that strategy games
> have used as a crutch.
>
> Anyway, here's the URL to the AI debugging session entitled "Human/Drengin
> wars". I through in flavor text as well to make it more interesting to
> read.
>
> http://www.galciv2.com/Journals.aspx?AID=98513

Looks interesting. I still suspect that computers will not play
strategy games at strong human level until they develop near human
level intelligence. Cheats or handicapping of some kind will always be
needed.

PS

- Gerry Quinn

In article <yZudnZa7eZRE8UneRVn-og@comcast.com>, bwardell@stardock.com
says...
> In general, my view was that the AI worked pretty well but militarily was
> disappointed. The main issue was its lack of coordination in projecting
> military power into a given quadrant. The result was hat looked like the
> age-old "Death train" (That's where units uselessly throw themselves at a
> wall dying with no purpose). There simply wasn't enough analysis on putting
> units together and using them effectivley.

An invariable problem, I suspect. You need strongish AI to strike a
balance between opportunistic raiding/exploration/dispersed defence and
attacks in force.

Part of the solution is often to change the rules. Civ4 did away with
separate attack and defence parameters, and instantly the cost to the
AI of ill-judged cavalry raids was reduced, though not eliminated.

> A lot -- A LOT -- of work has been going into the AI of Galactic
> Civilizations II in order to not just make it "challenging" but to make it
> play intelligently. There were significant chunks of the GalCiv I AI that
> were scripted. That is, the AI is programmed to do certain things due to
> play-testing over and over. This time around, we're trying our best to make
> the AI build strategies based on analysis and not rely on scripts.
>
> The improvements in the AI have led to a lot of changes, probably the
> biggest one being the elimination of the infamous "The AI knows where the
> good planets are" issue that GalCiv from the OS/2 days on up to the 2003
> release of GalCiv. GalCiv II's AI has to build scouts and look for good
> planets just like everyone else.

That was more of an issue because some people were claiming that the AI
played by exactly the same rules, I suspect ;-)

Of course, some people just don't understand the difficulty of writing
an AI and will always complain when sleight of hand is used to create
an effective computer opponent.

> To get an idea of what is involved in debugging such an AI, I chronicled one
> such game. Mind you, we're now doing several of these types of games each
> day to tune the AI, to find flaws, etc. The idea being that when the game's
> released, the difficulty levels will determine how smartly the AI plays
> rather than age old "Here's some money (or less of)" way that strategy games
> have used as a crutch.
>
> Anyway, here's the URL to the AI debugging session entitled "Human/Drengin
> wars". I through in flavor text as well to make it more interesting to
> read.
>
> http://www.galciv2.com/Journals.aspx?AID=98513

Looks interesting. I still suspect that computers will not play
strategy games at strong human level until they develop near human
level intelligence. Cheats or handicapping of some kind will always be
needed.

PS

- Gerry Quinn

In article <yZudnZa7eZRE8UneRVn-og@comcast.com>, bwardell@stardock.com
says...
> In general, my view was that the AI worked pretty well but militarily was
> disappointed. The main issue was its lack of coordination in projecting
> military power into a given quadrant. The result was hat looked like the
> age-old "Death train" (That's where units uselessly throw themselves at a
> wall dying with no purpose). There simply wasn't enough analysis on putting
> units together and using them effectivley.

An invariable problem, I suspect. You need strongish AI to strike a
balance between opportunistic raiding/exploration/dispersed defence and
attacks in force.

Part of the solution is often to change the rules. Civ4 did away with
separate attack and defence parameters, and instantly the cost to the
AI of ill-judged cavalry raids was reduced, though not eliminated.

> A lot -- A LOT -- of work has been going into the AI of Galactic
> Civilizations II in order to not just make it "challenging" but to make it
> play intelligently. There were significant chunks of the GalCiv I AI that
> were scripted. That is, the AI is programmed to do certain things due to
> play-testing over and over. This time around, we're trying our best to make
> the AI build strategies based on analysis and not rely on scripts.
>
> The improvements in the AI have led to a lot of changes, probably the
> biggest one being the elimination of the infamous "The AI knows where the
> good planets are" issue that GalCiv from the OS/2 days on up to the 2003
> release of GalCiv. GalCiv II's AI has to build scouts and look for good
> planets just like everyone else.

That was more of an issue because some people were claiming that the AI
played by exactly the same rules, I suspect ;-)

Of course, some people just don't understand the difficulty of writing
an AI and will always complain when sleight of hand is used to create
an effective computer opponent.

> To get an idea of what is involved in debugging such an AI, I chronicled one
> such game. Mind you, we're now doing several of these types of games each
> day to tune the AI, to find flaws, etc. The idea being that when the game's
> released, the difficulty levels will determine how smartly the AI plays
> rather than age old "Here's some money (or less of)" way that strategy games
> have used as a crutch.
>
> Anyway, here's the URL to the AI debugging session entitled "Human/Drengin
> wars". I through in flavor text as well to make it more interesting to
> read.
>
> http://www.galciv2.com/Journals.aspx?AID=98513

Looks interesting. I still suspect that computers will not play
strategy games at strong human level until they develop near human
level intelligence. Cheats or handicapping of some kind will always be
needed.

PS

- Gerry Quinn

"Brad Wardell" <bwardell@stardock.com> wrote in message
news:yZudnZa7eZRE8UneRVn-og@comcast.com...
> Back a couple of years ago there was a great deal of discussion on Usenet
> with regards to the GalCiv I AI. Most people thought it was pretty good
> but others felt the military tactics were bad. That while the AI provided
> a challenge, it mostly was able to do so because it was more effective at
> managing its economy/industry than the typical human and thus could
> essentially "win" through brute force.
>
> In general, my view was that the AI worked pretty well but militarily was
> disappointed. The main issue was its lack of coordination in projecting
> military power into a given quadrant. The result was hat looked like the
> age-old "Death train" (That's where units uselessly throw themselves at a
> wall dying with no purpose). There simply wasn't enough analysis on
> putting units together and using them effectivley.
>
> A lot of this came down to being restricted to CPU light algorithms.
> GalCiv I was a sprite-based game -- no hardware acceleration. So the
> graphics were CPU bound. Even though the AI is, and continues to be,
> multithreaded, there are still very real limits on how much time can be
> spent generating multi-turn strategies. GalCiv II, however, is
> 3D-accelerated which has freed up the CPU a great deal. This has allowed
> for greatly expanded algorithms during military endeavors without raising
> the hardware requirements (in fact, on the typical PC, GalCiv II will run
> better than GalCiv I did ironically despite immense graphics improvements
> and 3D engine).
>
> A lot -- A LOT -- of work has been going into the AI of Galactic
> Civilizations II in order to not just make it "challenging" but to make it
> play intelligently. There were significant chunks of the GalCiv I AI that
> were scripted. That is, the AI is programmed to do certain things due to
> play-testing over and over. This time around, we're trying our best to
> make the AI build strategies based on analysis and not rely on scripts.
>
> The improvements in the AI have led to a lot of changes, probably the
> biggest one being the elimination of the infamous "The AI knows where the
> good planets are" issue that GalCiv from the OS/2 days on up to the 2003
> release of GalCiv. GalCiv II's AI has to build scouts and look for good
> planets just like everyone else.
>
> To get an idea of what is involved in debugging such an AI, I chronicled
> one such game. Mind you, we're now doing several of these types of games
> each day to tune the AI, to find flaws, etc. The idea being that when the
> game's released, the difficulty levels will determine how smartly the AI
> plays rather than age old "Here's some money (or less of)" way that
> strategy games have used as a crutch.
>
> Anyway, here's the URL to the AI debugging session entitled "Human/Drengin
> wars". I through in flavor text as well to make it more interesting to
> read.
>
> http://www.galciv2.com/Journals.aspx?AID=98513
>
>

Very interesting. I am glad to see that the computer AI is being taken
seriously. For a few years, I was in the camp of thinking that multiplayer
was the cure for bad computer AI.

But now that I'm busier with life(TM) I prefer to play my strategy games
against the computer only. I just don't have time to spend a day looking for
someone else to play. So a smart and challenging computer player is very
important.

-Craig

"Brad Wardell" <bwardell@stardock.com> wrote in message
news:yZudnZa7eZRE8UneRVn-og@comcast.com...
> Back a couple of years ago there was a great deal of discussion on Usenet
> with regards to the GalCiv I AI. Most people thought it was pretty good
> but others felt the military tactics were bad. That while the AI provided
> a challenge, it mostly was able to do so because it was more effective at
> managing its economy/industry than the typical human and thus could
> essentially "win" through brute force.
>
> In general, my view was that the AI worked pretty well but militarily was
> disappointed. The main issue was its lack of coordination in projecting
> military power into a given quadrant. The result was hat looked like the
> age-old "Death train" (That's where units uselessly throw themselves at a
> wall dying with no purpose). There simply wasn't enough analysis on
> putting units together and using them effectivley.
>
> A lot of this came down to being restricted to CPU light algorithms.
> GalCiv I was a sprite-based game -- no hardware acceleration. So the
> graphics were CPU bound. Even though the AI is, and continues to be,
> multithreaded, there are still very real limits on how much time can be
> spent generating multi-turn strategies. GalCiv II, however, is
> 3D-accelerated which has freed up the CPU a great deal. This has allowed
> for greatly expanded algorithms during military endeavors without raising
> the hardware requirements (in fact, on the typical PC, GalCiv II will run
> better than GalCiv I did ironically despite immense graphics improvements
> and 3D engine).
>
> A lot -- A LOT -- of work has been going into the AI of Galactic
> Civilizations II in order to not just make it "challenging" but to make it
> play intelligently. There were significant chunks of the GalCiv I AI that
> were scripted. That is, the AI is programmed to do certain things due to
> play-testing over and over. This time around, we're trying our best to
> make the AI build strategies based on analysis and not rely on scripts.
>
> The improvements in the AI have led to a lot of changes, probably the
> biggest one being the elimination of the infamous "The AI knows where the
> good planets are" issue that GalCiv from the OS/2 days on up to the 2003
> release of GalCiv. GalCiv II's AI has to build scouts and look for good
> planets just like everyone else.
>
> To get an idea of what is involved in debugging such an AI, I chronicled
> one such game. Mind you, we're now doing several of these types of games
> each day to tune the AI, to find flaws, etc. The idea being that when the
> game's released, the difficulty levels will determine how smartly the AI
> plays rather than age old "Here's some money (or less of)" way that
> strategy games have used as a crutch.
>
> Anyway, here's the URL to the AI debugging session entitled "Human/Drengin
> wars". I through in flavor text as well to make it more interesting to
> read.
>
> http://www.galciv2.com/Journals.aspx?AID=98513
>
>

Very interesting. I am glad to see that the computer AI is being taken
seriously. For a few years, I was in the camp of thinking that multiplayer
was the cure for bad computer AI.

But now that I'm busier with life(TM) I prefer to play my strategy games
against the computer only. I just don't have time to spend a day looking for
someone else to play. So a smart and challenging computer player is very
important.

-Craig

"Brad Wardell" <bwardell@stardock.com> wrote in message
news:yZudnZa7eZRE8UneRVn-og@comcast.com...
> Back a couple of years ago there was a great deal of discussion on Usenet
> with regards to the GalCiv I AI. Most people thought it was pretty good
> but others felt the military tactics were bad. That while the AI provided
> a challenge, it mostly was able to do so because it was more effective at
> managing its economy/industry than the typical human and thus could
> essentially "win" through brute force.
>
> In general, my view was that the AI worked pretty well but militarily was
> disappointed. The main issue was its lack of coordination in projecting
> military power into a given quadrant. The result was hat looked like the
> age-old "Death train" (That's where units uselessly throw themselves at a
> wall dying with no purpose). There simply wasn't enough analysis on
> putting units together and using them effectivley.
>
> A lot of this came down to being restricted to CPU light algorithms.
> GalCiv I was a sprite-based game -- no hardware acceleration. So the
> graphics were CPU bound. Even though the AI is, and continues to be,
> multithreaded, there are still very real limits on how much time can be
> spent generating multi-turn strategies. GalCiv II, however, is
> 3D-accelerated which has freed up the CPU a great deal. This has allowed
> for greatly expanded algorithms during military endeavors without raising
> the hardware requirements (in fact, on the typical PC, GalCiv II will run
> better than GalCiv I did ironically despite immense graphics improvements
> and 3D engine).
>
> A lot -- A LOT -- of work has been going into the AI of Galactic
> Civilizations II in order to not just make it "challenging" but to make it
> play intelligently. There were significant chunks of the GalCiv I AI that
> were scripted. That is, the AI is programmed to do certain things due to
> play-testing over and over. This time around, we're trying our best to
> make the AI build strategies based on analysis and not rely on scripts.
>
> The improvements in the AI have led to a lot of changes, probably the
> biggest one being the elimination of the infamous "The AI knows where the
> good planets are" issue that GalCiv from the OS/2 days on up to the 2003
> release of GalCiv. GalCiv II's AI has to build scouts and look for good
> planets just like everyone else.
>
> To get an idea of what is involved in debugging such an AI, I chronicled
> one such game. Mind you, we're now doing several of these types of games
> each day to tune the AI, to find flaws, etc. The idea being that when the
> game's released, the difficulty levels will determine how smartly the AI
> plays rather than age old "Here's some money (or less of)" way that
> strategy games have used as a crutch.
>
> Anyway, here's the URL to the AI debugging session entitled "Human/Drengin
> wars". I through in flavor text as well to make it more interesting to
> read.
>
> http://www.galciv2.com/Journals.aspx?AID=98513
>
>

Very interesting. I am glad to see that the computer AI is being taken
seriously. For a few years, I was in the camp of thinking that multiplayer
was the cure for bad computer AI.

But now that I'm busier with life(TM) I prefer to play my strategy games
against the computer only. I just don't have time to spend a day looking for
someone else to play. So a smart and challenging computer player is very
important.

-Craig