Matrix Games Forums

(sorry for bad english)
I'v didn't not visit here for more than 6 months. I'm happy with the news about AE and the improvement is looks great.

I think most important feature would be strategy AI. You know, This game is a some kind of chess game.

when i played original WITP as a japan side, AI ABCD fleet always waste ther precious warships by sending to hopeless bataan.

ABCD fleet should defend DEI area.

IMHO Good AI means

1. realize a situation
2. make logical plan
3. historical restriction

I wonder how is going aout strategy AI.

I think Rommel3 meant ABDA ...

ORIGINAL: Rommel3

(sorry for bad english)
I'v didn't not visit here for more than 6 months. I'm happy with the news about AE and the improvement is looks great.

I think most important feature would be strategy AI. You know, This game is a some kind of chess game.

when i played original WITP as a japan side, AI ABCD fleet always waste ther precious warships by sending to hopeless bataan.

ABCD fleet should defend DEI area.

IMHO Good AI means

1. realize a situation
2. make logical plan
3. historical restriction

I wonder how is going aout strategy AI.

That's hard to do. We're looking at biologically-inspired AI for autonomous vehicle control, and there are some very recent results from neuroscience that are beginning to show us the direction to go. Perhaps an AI for a game would be a good demonstration.

ORIGINAL: Rommel3

(sorry for bad english)
I'v didn't not visit here for more than 6 months. I'm happy with the news about AE and the improvement is looks great.

I think most important feature would be strategy AI. You know, This game is a some kind of chess game.

when i played original WITP as a japan side, AI ABCD fleet always waste ther precious warships by sending to hopeless bataan.

ABCD fleet should defend DEI area.

IMHO Good AI means

1. realize a situation
2. make logical plan
3. historical restriction

I wonder how is going aout strategy AI.

The only "guarantee" is that it will be "at least as good as the current AI". The kind of re-write you are hoping for is probably WITP II at the earliest.

AE team members repeatedly stated that they are working to improve AI components. They were careful not to set any expectations. But there is hope ...

ORIGINAL: Barb

I think Rommel3 meant ABDA ...

ABCD is another name for a similar grouping ....


A=America
B=Britain
C=China
D=Dutch

ORIGINAL: jwilkerson

ORIGINAL: Barb

I think Rommel3 meant ABDA ...

ABCD is another name for a similar grouping ....


A=America
B=Britain
C=China
D=Dutch

next time won't you sing with me....[;)]

That's hard to do. We're looking at biologically-inspired AI for autonomous vehicle control, and there are some very recent results from neuroscience that are beginning to show us the direction to go. Perhaps an AI for a game would be a good demonstration.

Are you using the Machine Learning method like the chess software?

ORIGINAL: Saso

That's hard to do. We're looking at biologically-inspired AI for autonomous vehicle control, and there are some very recent results from neuroscience that are beginning to show us the direction to go. Perhaps an AI for a game would be a good demonstration.

Are you using the Machine Learning method like the chess software?

Yes, we think learning is important to intelligence--and we have machine learning experts on staff--but there are other things that need to be done as well. There's a recent paper in J Neuroscience (O'Reilly, et al., 2008, J Neurosci, 28(9):2252-2260, Feb 27, 2008) that provides insight into the role of the cerebellum in learning forward and backward models, and those models are needed for goal-directed behaviour (behaviour where a change in the final reward propagates backwards to the current choice of actions). Chess software gives some insight, but optimising behaviour is relative to a continuous manifold of possible plans.

ORIGINAL: herwin

ORIGINAL: Saso

That's hard to do. We're looking at biologically-inspired AI for autonomous vehicle control, and there are some very recent results from neuroscience that are beginning to show us the direction to go. Perhaps an AI for a game would be a good demonstration.

Are you using the Machine Learning method like the chess software?

Yes, we think learning is important to intelligence--and we have machine learning experts on staff--but there are other things that need to be done as well. There's a recent paper in J Neuroscience (O'Reilly, et al., 2008, J Neurosci, 28(9):2252-2260, Feb 27, 2008) that provides insight into the role of the cerebellum in learning forward and backward models, and those models are needed for goal-directed behaviour (behaviour where a change in the final reward propagates backwards to the current choice of actions). Chess software gives some insight, but optimising behaviour is relative to a continuous manifold of possible plans.

I speak English....and I have no idea what you just said. [:D]

ORIGINAL: Charbroiled

ORIGINAL: herwin

ORIGINAL: Saso



Are you using the Machine Learning method like the chess software?

Yes, we think learning is important to intelligence--and we have machine learning experts on staff--but there are other things that need to be done as well. There's a recent paper in J Neuroscience (O'Reilly, et al., 2008, J Neurosci, 28(9):2252-2260, Feb 27, 2008) that provides insight into the role of the cerebellum in learning forward and backward models, and those models are needed for goal-directed behaviour (behaviour where a change in the final reward propagates backwards to the current choice of actions). Chess software gives some insight, but optimising behaviour is relative to a continuous manifold of possible plans.

I speak English....and I have no idea what you just said. [:D]

Yes, learning is part of being smart, but there are other things involved, too. The piece of the brain at the back seems to be used to look into the future and back from the future to the now. Being able to do that is part of planning. We just don't know how it does it. Chess software does it, too, but animals seem to be able to deal with a continuous world instead of one involving single moves.

Chess software does it, too, but animals seem to be able to deal with a continuous world instead of one involving single moves.

The problem with chess software is that it learns in a "monotonous" way?
I mean, each move that it makes is in relation with a determined value, therefore it's very capable to analyze the chess endgame positions but it's a lack of depth to analyze the strategic positions (of course relatively speaking because they are more strong than the 99% of the players).

I'm just asking because I'd like to known how this system works. [:)]

ORIGINAL: Saso

Chess software does it, too, but animals seem to be able to deal with a continuous world instead of one involving single moves.

The problem with chess software is that it learns in a "monotonous" way?
I mean, each move that it makes is in relation with a determined value, therefore it's very capable to analyze the chess endgame positions but it's a lack of depth to analyze the strategic positions (of course relatively speaking because they are more strong than the 99% of the players).

I'm just asking because I'd like to known how this system works. [:)]

Basically, goal-directed behaviour involves working out a collection of action plans to reach a reward and then propagating the reward back to get values for the actions in the plan. Finally, you compare the values for various actions currently available to choose the best. There's a lot more, since you also have reflexive and habitual behaviour and you have to take into account uncertainty. The brain seems to use at least three subsystems in handling goal-oriented behaviour--the dorso-lateral-prefrontal cortex chooses actions, the basal ganglia come up with reward estimates, and the posterior cerebellum propagates actions forward in time to predict the state after each action and backward in time to associate values with actions. The connectivity is very foggy, but I think I know how to model each subsystem.

I did some searching in the recent literature--in large part because we're planning a grant proposal for an autonomous robot. A biologically-inspired AI for a strategy game would consist of four major (distributed) modules:
1. Memory for the current situation.
2. Memory for the planned situation currently being evaluated.
3. A predictor module that moves the plan one step forward (on a broad front, using a large number of possible moves).
4. An evaluator module that checks the planned situation after the step forward and generates error signals for all the bad results.
The error signals would cause all the bad moves at the previous step to be unlearned. There would be thousands of error signals, up to ten cells per error signal, and each cell would monitor up to 150,000 numerical state descriptors. The initial weights would be set based on alpha testing, and then polished over time so that an AI would gradually learn how to thrash its usual opponent.

Even a grossly simplified version of this would probably be a challenge for a newbie.

Yes, thanks for your explanation [:)]