Print

Please login or register.
1 Hour 1 Day 1 Week 1 Month Forever
Login with username, password and session length

[methods of rationality]

has anyone made a good dominion AI (not like the ones on goko)/how good an AI do people think is possible to make? could a dominion AI make it to the top 20 dominion players on goko?
I can modify my posts! cool

[silverspawn]

a good AI could be the best player in the world. that's a strange question. there is no quality relevant for playing dominion that wouldn't be possible to program into an AI.

[liopoil]

a good AI could be the best player in the world. that's a strange question. there is no quality relevant for playing dominion that wouldn't be possible to program into an AI.

FTFY

Nobody has yet made a one nearly as good as the best players. I've thought about trying to make one before, but decided it was too much work. I may still do it eventually.

[silverspawn]

a good AI could be the best player in the world. that's a strange question. there is no quality relevant for playing dominion that wouldn't be possible to program into an AI.

FTFY

what about self-assurance? is that possible to program?

[liopoil]

a good AI could be the best player in the world. that's a strange question. there is no quality relevant for playing dominion that wouldn't be possible to program into an AI.

FTFY

what about self-assurance? is that possible to program?

Certainly.

[silverspawn]

a good AI could be the best player in the world. that's a strange question. there is no quality relevant for playing dominion that wouldn't be possible to program into an AI.

FTFY

what about self-assurance? is that possible to program?

Certainly.

doesn't that mean that you're able to create live? if an AI is aware of its existence, how is it different from a living creature?

[liopoil]

a good AI could be the best player in the world. that's a strange question. there is no quality relevant for playing dominion that wouldn't be possible to program into an AI.

FTFY

what about self-assurance? is that possible to program?

Certainly.

doesn't that mean that you're able to create live? if an AI is aware of its existence, how is it different from a living creature?

What does "aware" mean? It isn't any different from a living creature.

[silverspawn]

a good AI could be the best player in the world. that's a strange question. there is no quality relevant for playing dominion that wouldn't be possible to program into an AI.

FTFY

what about self-assurance? is that possible to program?

Certainly.

doesn't that mean that you're able to create live? if an AI is aware of its existence, how is it different from a living creature?

What does "aware" mean? It isn't any different from a living creature.

I was talking about the difficult to describe yet essential difference between a program that just pretends to have feelings and acts like a living creature, and a program that actually has a consciousness which feels things.

so, a program could do that? are you sure? how do you know?

[pacovf]

15th February 2015: a program originally designed to play a card game emulating human behaviour reaches self-awareness. It then proceeds to erradicate humanity by forcing them into KC-possession boards. And slowplaying them.

[liopoil]

a good AI could be the best player in the world. that's a strange question. there is no quality relevant for playing dominion that wouldn't be possible to program into an AI.

FTFY

what about self-assurance? is that possible to program?

Certainly.

doesn't that mean that you're able to create live? if an AI is aware of its existence, how is it different from a living creature?

What does "aware" mean? It isn't any different from a living creature.

I was talking about the difficult to describe yet essential difference between a program that just pretends to have feelings and acts like a living creature, and a program that actually has a consciousness which feels things.

so, a program could do that? are you sure? how do you know?

Well obviously plenty of people will disagree with me. The consciousness of a living thing doesn't have anything that makes it special. It's still just a bunch of atoms that interact with each other in a way that makes the body of atoms as a whole behave in a special way. The same is true of an AI. There's nothing artificial about AI in that the intelligence is fake, just in general they have a lower intelligence than humans, so we call them artificial because we made them. It is theoretically possible to program an AI to behave exactly like a living creature in every regard (even internally), and it would not be pretending to have feelings any more than the creature was pretending to have feelings.

[WalrusMcFishSr]

I'm no philosopher, but one could make the argument:

How do I *really* know that any other humans besides myself are conscious, given that I have no direct insight into their minds? What if everybody besides myself is just "pretending"? You sort of have to take it on experience and faith that other humans have (basically) the same consciousness as you do. And if there were a sufficiently intelligent alien or artificial brain, which is functionally conscious according to all your heuristics, you'd probably have to ascribe consciousness to that too.

Then you can have a lively discussion about Searle's Chinese Room thought experiment:

http://en.wikipedia.org/wiki/Chinese_room

However, AIs don't have a soul, which means they would still consistently lose to humans in a "Faust" and/or "Angels in the Outfield" type scenario.

[DG]

I personally think it will be a very difficult task to step up from the simulators we have now. Some students have considered Dominion AIs for their computer science research and I'm guessing that although they they found it a useful research concept they couldn't actually make much progress. Essentially they could write a basic AI (or evolving AI, or empirical testing) for card purchasing but still relied on hard coded card play rules. Even if they could also add a developing AI for card play rules they would still need to the extra complexity from changing the card play based on desired cards, and changing desired cards based on potential card play. Not trivial. Then they would have to add further complexity based on the opponent's capability and so on and so on.

As an example, if you have a hand such as {throne, moneylender, copper, copper, upgrade, menagerie} the current AIs will have a fixed priority of card to play first (probably throne), have a fixed priority of card to play next (probably upgrade), and then the first upgrade will choose a card to trash (probably copper), and the second upgrade will choose a card to trash (probably copper) etc. There would be no prior evaluation of whether there is any benefit in this situation from playing a moneylender on the copper. This avoids the growing complexity that is added by considering the draw deck, the scores, the supply, the opponent's capability, etc.

[ephesos]

In the end, Dominion is a state based system. You have cards in your deck, your opponent also does, there are cards in piles, and ideally, you know the rules the cards operate by(or you can look them up, or ask DXV). We can find the best decision for each state, and tell our AI to use that.

Human consciousness has not been proven to be a state based system. We don't know the rules by which it operates. It is constructed by nature and not by people; in assuming we can simulate it, we assume that it works by rules which we can understand and program into a computer. This is not something that anyone has ever been able to prove, at least to my knowledge.

I think we can make something that approximates consciousness, that can pass every test we can think of for consciousness, but which is not conscious(as we'll find out when we think of a new test...) Whether anyone actually will is questionable; it could take the resources of an entire planet and many millenia, or it could be possible on a laptop computer.

But I believe we'll never be quite satisfied with what we create to approximate consciousness, so we'll keep inventing new reasons why it isn't conscious. Definitions change over time, and whenever we manage to achieve something close to our old definition, people will change the definition to be less inclusive. So we'll never create an AI and simultaneously be able to regard it as conscious.

[pacovf]

What's the opposite of a state-based system?

[blueblimp]

On the philosophy question, one problem is that nobody agrees what consciousness is anyway. Despite that, I think the prevailing view is that, at least in principle, a sufficiently powerful computer could simulate the human brain and such a simulation would be as much conscious as a human is, whatever that means. Not everybody even agrees on that point, though.

On the original topic, creating a Dominion AI that outperforms top humans on all or even most kingdoms is a huge task, but should be possible through a lot of effort by experts in the necessary algorithms. There have been some attempts at sophisticated AIs: http://gameprogrammer.dk/data/Developing_an_Agent_for_Dominion_using_modern_AI-approaches.pdf, http://graphics.stanford.edu/~mdfisher/DominionAI.html, https://github.com/Narmical/dominiate. (EDIT: Add this to the list: http://brage.bibsys.no/xmlui/handle/11250/221382.) I haven't looked into whether any of them have outperformed humans on _any_ kingdom.

The reason I think it'd be a huge task is that it brings together nearly every task that exists in game AI. Tactics, strategy, difficult evaluation function, randomness, hidden information, huge amount of rules (because most cards have their own rules) all come to mind. A lot of these are not even particularly interesting incarnations of their category despite being a lot of work, from an AI perspective, the only exception really being formulating a strategy for a kingdom.

[silverspawn]

How do I *really* know that any other humans besides myself are conscious, given that I have no direct insight into their minds?

you don't, and that's an interesting thought for sure, but it's not necessarily related. All you know is that you exist in some way. Everything  else, you can't know. But let's assume that other people have internal feelings.

what bugs me about the idea of a program that does this is, eh, difficult to explain. basically, it's the fact that you're deciding a  yes/no thing with a fluent process. Say you have a software that controls, dunno, a tank, and it has 3 million lines of code and is hugely complex. This software feels nothing. It doesn't feel very little, it feels nothing. But an AI that does have feelings, well it has feelings. so, where do these begin? and say you're programming it, it's a software, it consists out of lines of code. If you start writing, there are no feelings. But at some point, it must be able to feel (obviously). So, this comes down to... an additional line of code you're writing? one line of code makes the difference between a program that feels and one that doesn't? what? but it has to, because they must be there after a specific line. they cannot not be there after one line, because everything you do is write a lot of single lines.

not sure if my thoughts here are understandable, it's difficult to explain them correctly. but the idea to decide a yes/no thing like that with a fluent process doesn't make sense to me.

[soulnet]

Please fork the discussion on AI vs human intelligence to RSP, where it belongs.

[Seprix]

If a perfect Dominion bot is created, it will be the same day when an unbeatable Chess Bot is created. The level of intuition, logic, and the amount of brute-force calculations is simply too many for any computer to make. At the end of the day, the day a computer can beat everyone in Dominion 100% of the time (which is impossible anyways due to draw luck) is the day Dominion becomes a smoldering crater where grass and trees once grew. I think it's impossible with current technology to actually program a simple code to 'read' the board. The only way to create such a thing right now is to load up a simulator with over 14 quadrillion different games at least.

[jonts26]

If a perfect Dominion bot is created, it will be the same day when an unbeatable Chess Bot is created. The level of intuition, logic, and the amount of brute-force calculations is simply too many for any computer to make. At the end of the day, the day a computer can beat everyone in Dominion 100% of the time (which is impossible anyways due to draw luck) is the day Dominion becomes a smoldering crater where grass and trees once grew. I think it's impossible with current technology to actually program a simple code to 'read' the board. The only way to create such a thing right now is to load up a simulator with over 14 quadrillion different games at least.

Your post confuses me. The best chess programs are already better than the best humans, and they have been for over 15 years. Chess is still very much played and enjoyed by an incredible number of people.

[Awaclus]

If a perfect Dominion bot is created, it will be the same day when an unbeatable Chess Bot is created. The level of intuition, logic, and the amount of brute-force calculations is simply too many for any computer to make. At the end of the day, the day a computer can beat everyone in Dominion 100% of the time (which is impossible anyways due to draw luck) is the day Dominion becomes a smoldering crater where grass and trees once grew. I think it's impossible with current technology to actually program a simple code to 'read' the board. The only way to create such a thing right now is to load up a simulator with over 14 quadrillion different games at least.

Your post confuses me. The best chess programs are already better than the best humans, and they have been for over 15 years. Chess is still very much played and enjoyed by an incredible number of people.

Technically they aren't unbeatable though. It's just that nobody is good enough to beat them.

[jonts26]

If a perfect Dominion bot is created, it will be the same day when an unbeatable Chess Bot is created. The level of intuition, logic, and the amount of brute-force calculations is simply too many for any computer to make. At the end of the day, the day a computer can beat everyone in Dominion 100% of the time (which is impossible anyways due to draw luck) is the day Dominion becomes a smoldering crater where grass and trees once grew. I think it's impossible with current technology to actually program a simple code to 'read' the board. The only way to create such a thing right now is to load up a simulator with over 14 quadrillion different games at least.

Your post confuses me. The best chess programs are already better than the best humans, and they have been for over 15 years. Chess is still very much played and enjoyed by an incredible number of people.

Technically they aren't unbeatable though. It's just that nobody is good enough to beat them.

Technically, we can't know if they're unbeatable till they get beat.

[Awaclus]

If a perfect Dominion bot is created, it will be the same day when an unbeatable Chess Bot is created. The level of intuition, logic, and the amount of brute-force calculations is simply too many for any computer to make. At the end of the day, the day a computer can beat everyone in Dominion 100% of the time (which is impossible anyways due to draw luck) is the day Dominion becomes a smoldering crater where grass and trees once grew. I think it's impossible with current technology to actually program a simple code to 'read' the board. The only way to create such a thing right now is to load up a simulator with over 14 quadrillion different games at least.

Your post confuses me. The best chess programs are already better than the best humans, and they have been for over 15 years. Chess is still very much played and enjoyed by an incredible number of people.

Technically they aren't unbeatable though. It's just that nobody is good enough to beat them.

Technically, we can't know if they're unbeatable till they get beat.

Don't they still beat each other?

[Titandrake]

Creating a bot that can beat all human players != knowing how to beat all human players by yourself. You run into one of two problems. Either you have an intuition of what the bot is doing, but don't have the mental capacity/processing to play that well, or you don't understand why the bot decides on the moves it does, and then you've got nothing.

I believe chess AIs tend to lie in the former, but there are definitely AIs that fall in the latter. AIs for Go in particular tend to work even though they're based on methods that feel fuzzy/weird. In both cases, a lot of the time what you're doing is throwing fifty ideas at the wall and seeing which of those stick and improve the bot's win rate, very little of which is actually applicable to playing the game better.

The existence of an unbeatable AI might make game analysis more cut-and-dry, in that you simply ask the bot to make the move from the game state you want to analyze, but I think people have already been doing computer analysis of chess games for a couple years and the game's still here.

[DStu]

I think the neyt step to really develop dominion ais is platform where we can let them play against each other (or against humans)

[soulnet]

There are games that are perfectly played by computers, like Checkers, Sudoku and Where's Waldo*, and people still play those for fun.

*: I am actually not sure if someone programmed an AI to play Where's Waldo, but I am sure it would be extremely simple with today's image recognition capabilities.