Blake: On BtS AI Programming...?

[Proteus_MST]

Originally posted by Unimatrix11
I am not a programmer at all and i also mentioned this like two years ago. Also this is not to insult anyone or complain to anyone in specfic, BUT: I just wonder how on earth this chess-program on my C64 ("Grandmaster"), consisting of 8 KBytes of code (graphics included) and running on a 0.97 Mhz 8-bit-processor, managed to beat my ass pretty quickly (1 or 2 minutes per move IIRC) every single time for years... i actually never managed to beat it - and it didnt get any bonuses AFAIK ...

Yeahyeah - i know: It´s probably because i am a really poor chess player - but i am not (tho not a exactly a "grand-master" neither)...

A good example about the difference that the number of possible moves makes for an AI is if you compare Chess programs to programs for the game of Go (which Blake already mentioned).

Go has very few rules, much less than you have with chess, but this also means that you can place your stones almost everywhere on the board and can, especialy at the begnning, use several strategies that all can be considered as good moves (and there also are no difference in the value of the stones from the beginning on, their worth for your strategy is only created when you place them on the board).

As a result in contrast to chess programs that can beat grandmasters even the best Go programs can only beat Go players with the grade around 12. Kyu (like Karate Go uses a system of Kyu (pupil) and Dan (master) levels.)

[rjmatsleepers]

Originally posted by Unimatrix11
BTW: What is the factor of processing power when you compare a C64 with a total of 64KBytes of RAM with a 2 or 3 Ghz 32/64 bit processor, assisted by a modern graphics card and 1 or 2 GBytes of RAM (running on Windows XP / Vista) ? I know it´s pretty impossible to give a straight foward answer to that question, and it is rather rethorical anyways - the difference must be huge.

As you say, its not easy to give a simple answer. There is a table on Wikipedia giving millions of instructions per second (mips) for a number of computers over the last 30 years. This seems to suggest that the typical mid to late 80s micro processor had perhaps 0.5 mips compared with 20,000 mips from a modern chip. In other words an improvement of less than ten to the power five - much less than ten to the twelve, let alone ten to the twenty four.

RJM

[Maniac]

Originally posted by OliverFA
GalCiv has an option for the AI to be "thinking" always, including the player's turn. Would something like that make Civ AIs more powerful?

[joncnunn]

This looks like syematics, all possible moves with better pruning == a small number in the look ahead mode.

There's a great varance in human chess players on how far down they examine promising moves. And also the masters end up memorizing the start game & end game tables.

On the sacific thing, weather the AI sees it as a good move or not depends entirely upon weather the benifit is on this side of the horizon or not, and accordingly that was one of the last things for their AIs to get (via increased depth search + secondary search).
The related thing the horizon effect was noticed for was if a piece was doomed but it's loss could be delayed one more turn at a greater cost when it finally did fall, the AI would delay it.

Originally posted by rjmatsleepers

I was under the impression that research into human chess playing had concluded that humans do not evaluate all possible moves, rather, they consider only a small number. That only in complicated situations involving multiple captures do humans look many turns ahead. Their strength is (as you say) in their ability to evaluate a position. As a result, humans will miss obscure good moves such as a sacrifice that has a benefit many moves ahead.

RJM

[joncnunn]

Writing a good AI for Go is indeed a much harder problem.
I'm not even sure if brute force on todays supercomputer would complete in a reasonable timeframe.

Originally posted by Proteus_MST

A good example about the difference that the number of possible moves makes for an AI is if you compare Chess programs to programs for the game of Go (which Blake already mentioned).

Go has very few rules, much less than you have with chess, but this also means that you can place your stones almost everywhere on the board and can, especialy at the begnning, use several strategies that all can be considered as good moves (and there also are no difference in the value of the stones from the beginning on, their worth for your strategy is only created when you place them on the board).

As a result in contrast to chess programs that can beat grandmasters even the best Go programs can only beat Go players with the grade around 12. Kyu (like Karate Go uses a system of Kyu (pupil) and Dan (master) levels.)

[rjmatsleepers]

Originally posted by joncnunn
This looks like syematics, all possible moves with better pruning == a small number in the look ahead mode.

There's a great varance in human chess players on how far down they examine promising moves. And also the masters end up memorizing the start game & end game tables.

On the sacific thing, weather the AI sees it as a good move or not depends entirely upon weather the benifit is on this side of the horizon or not, and accordingly that was one of the last things for their AIs to get (via increased depth search + secondary search).
The related thing the horizon effect was noticed for was if a piece was doomed but it's loss could be delayed one more turn at a greater cost when it finally did fall, the AI would delay it.

I think there is a genuine difference between the way in which computers decide which move to make in a "brute force" chess playing program and the way in which a human does it. This goes beyond semantics. In the middle game, a "brute force" computer program will construct a look ahead tree for all possible moves. Alpha beta pruning and killer heuristics will reduce the actual computation required, but in principle all positions at a given depth are evaluated.

Humans don't do this. They will restrict themselves to moves which achieve a particular goal, or meet a particular threat. Some moves are not considered at all. It is not that they are rapidly rejected, they are simply not considered.

Attempts have been made to introduce chess knowledge and goal oriented behavour to chess playing programs, but my understanding is that these have generally not resulted in a higher standard of play.

Given the size of the look ahead tree and the random elements involved, I suspect that this would not be true of a Civ 4 playing program.

RJM

[Harrier UK]

Re: Re: Blake: On BtS AI Programming...?

Originally posted by vilemerchant


What makes you think the 'default' level is ever supposed to be any sort of actual challenge? The AI doesn't get insane bonuses on harder levels, it gets sensible bonuses that allow a computer to compete with more advanced human players. What you get on Noble is a level of play which applies to a n00b that just opened the box and doesn't really know what he's doing. It's easy because it's meant to be easy, there's 5 higher levels that apply to people who want it harder. Any lower than that and it's basically in 'baby mode' so your girlfriend/wife/any other female or AAHZ can play. :P

Just started to read this old thread - so I do not know what has been said after this post.

The default level is not Noble, Noble is the median level. Suppossably no bonuses or reductions. So the player or AI have no advantages.

Also your comment that players that play on the Noble level are "Noobs" is an insult to many thousands of players.