Want to Help Create a Better AI?

If you wish to help create a better AI for Civilization games, please read the following:

Dear All,
Arago, an artificial intelligence company based in Frankfurt and New York invites you to participate in our HIROic challenge. We are working on an artificial intelligence, called HIRO, and we believe that it will be able to beat the world’s best players in Freeciv shortly. However, HIRO doesn’t know anything about the game yet.

Although HIRO is a quick learner it needs awesome mentors: world class Civ experts that know how to build civilizations and compete against other empires within the game. Therefore, we are looking for the most experienced and most creative players in your community that are ready to support us to not only teach HIRO to play Freeciv like a world champion but to create the world’s most powerful computer opponent.

The major difference between the usual computer opponents you are already familiar with and HIRO is that the latter will be human-like – no more tricking, no more unfair information advantages. Every time you play against HIRO you will be playing against the world’s best players at the same time, because HIRO will pool the knowledge of the world’s best human players in one “super brain”.

The task we ask for is simple: play at least two games with us in the beginning of August and tell us about the principles of your pathway to victory. Once you register, you work with a "buddy" from our organization. We have set up a Docker container that will have everything you need to participate and we will assist you in getting setup to play. Any issues can be addressed through your "buddy".

Depending upon how well you do, your efforts will be rewarded with great non-cash bonuses. Furthermore, Arago would like to provide your community with its computer opponent as soon as it is ready to win against the best.

If you want to go down in history, please register at http://hiroic.arago.co and beat the beast.

Regards,
Arago Team

We are intentionally exempting this from our policy of no advertisements because we feel this effort could bring great benefit to the civilization community.

Challenge accepted.

Great initiative, Leif! Comes right on time in these days of decaying quality of AI in our franchise.

Good Luck!

Aristos said:

Challenge accepted.

Great initiative, Leif! Comes right on time in these days of decaying quality of AI in our franchise.

Click to expand...

Ahem!!

G

Gazebo said:

Ahem!!

G

Click to expand...

I know man, VP was obviously not included in stated decay. I was referring to the industry that collects money for their no-AI, not to the modders that save the product for free.

Sweet! Now I can live my life long dream of training the first generation of robotic overlords.

Awesome! I'd be game anyways, but a chance to win civ 6 makes it even better.

We'll start in the next days. To show you a proof, that we are already on it (and to give you a peek ahead) here is a short screencast:

https://drive.google.com/file/d/0B7wHcD_BNVY2Q2N0QWZGQ3AtV2M/view?invite=CMCo25sP

See you soon,
Joker

P.S.
If you have any questions, I'd be glad to help

Which version of Freeciv will the AI trained on?

Hans Lemurson said:

Which version of Freeciv will the AI trained on?

Click to expand...

Currently we are using 2.5.4.

Are you planning on sticking with one version for the whole run, or will you update periodically? And if so, how would you control for any changes to the game that might happen? The ruleset is mature enough that I don't foresee any major changes occurring, so it probably wouldn't matter anyways.

Also, I've read that in a number of cases, the native AIs for FreeCiv don't actually play by exactly the same ruleset as the player. I am assuming that HIRO will be playing on a level field with it's human trainers.

Speaking of which, what exactly will this challenge entail? Is HIRO in any state yet to win a full game? Is this a 1v1 duel against HIRO? 1 Human versus many HIROs? The website says that we are to "play two games". Will 2 matches per contestant be enough to get the data you need?

Given what's the current craze in machine learning, and given that you're collecting "truth", may I guess that the backbone involves deep learning?

Hans Lemurson said:

Are you planning on sticking with one version for the whole run, or will you update periodically?

Click to expand...

Most probably we will stick to this version, since there is no need for an update. The ruleset will not change and bugs are mainly in the AI (which we replace anyways )

Hans Lemurson said:

And if so, how would you control for any changes to the game that might happen?
The ruleset is mature enough that I don't foresee any major changes occurring, so it probably wouldn't matter anyways.

Click to expand...

There would be an impact only if the changes would affect the ruleset. This is very unlikely.

Hans Lemurson said:

Also, I've read that in a number of cases, the native AIs for FreeCiv don't actually play by exactly the same ruleset as the player. I am assuming that HIRO will be playing on a level field with it's human trainers.

Click to expand...

This is exactly what we are about to get rid of. The AI has more knowledge (i.e. it knows how well a city is defended) than the player and is stronger by "cheating".
HIRO will act as a human player with only the knowledge a player has and with the ruleset a player uses.

Hans Lemurson said:

Speaking of which, what exactly will this challenge entail?

Click to expand...

We contact each registered user from tomorrow on. The whole process is explained in detail there.
To give a short overview i have to explain a bit. HIRO is working with very small amounts of knowledge artifacts. Your buddy (a guy assigned to you) will extract your style of play and your tricks while watching you play. This will build a "knowledge base" used by HIRO. Each player has a knowledge base and those will challenge each other.

Hans Lemurson said:

Is HIRO in any state yet to win a full game?

Click to expand...

To be honest HIRO is capable of loosing a game
The rules enable it to do legal turns and to follow a strategy. But in the moment it plays worse than I do - which is already poor.

Hans Lemurson said:

Is this a 1v1 duel against HIRO? 1 Human versus many HIROs?

Click to expand...

The engine is free to use for every one participating. You can play with you friends, with normal KI and with HIRO in any combination you like to. For the competition we'll use the trained HIRO for every player to keep it fair and comparable.

Hans Lemurson said:

The website says that we are to "play two games". Will 2 matches per contestant be enough to get the data you need?

Click to expand...

You are welcome to play more than those, but we don't want to annoy you more than agreed. The mentioned two games are those to be played with your buddy in order to extract your style of play and your tricks. I assume that in two games the most situations occure and that additional ideas will come up without another game.

The_J said:

Given what's the current craze in machine learning, and given that you're collecting "truth", may I guess that the backbone involves deep learning?

Click to expand...

Not really. I'll explain the idea of HIRO in a very abstract and simplified way:
HIRO acts like a human. Under a given situation he is using his knowledge to find the possibilities which are applicable and selects the one which will bring the best results.

For example if you have a settler, you can move it to a better position, build a city or merge it with an existing city. three possible actions. Depending on the situation each action can be useful. As a human you will act "intuitively", which means your brain is judging the situation, combines it with your strategy and you experience. HIRO will do the same.

Deep learning (as in Alpha GO) means you take a look at billions of matches and extract a strategy out of this. This has some big disadvantages:
It is slow as hell, takes tons of CPU power and is rarely suitable for uncertain situations. Well go a different way, giving the human ideas more weight. The learning takes place bey judging all the giving possibilities. We call this experience.

Once again: This is a very simplified explanation. If somebody is interested in the real process I'd be happy to explain it. It will take some hours and probably a skype session.

No need, thanks , was just interested if this was an approach with deep learning, since it's the rage at the moment and probably the only form of machine learning which anyone of the general public might know.
Would still be curious though about how you're implementing, if you consider supervised learning not an option (although I don't see the problem with the computing power, because I'd guess that it'd would be only necessary to train it once after enough ground truth has been collected; and there's the cloud....; but then again, personally only experience with unsupervised learning and smaller datasets, no idea how this scales up). Would be satisfied with 1-2 sentences though ^^.


On the practical side: Don't think I have time to play games at the moment, my computer is a tad bit too old, so probably not gonna happen.

The_J said:

No need, thanks , was just interested if this was an approach with deep learning, since it's the rage at the moment and probably the only form of machine learning which anyone of the general public might know.
Would still be curious though about how you're implementing, if you consider supervised learning not an option (although I don't see the problem with the computing power, because I'd guess that it'd would be only necessary to train it once after enough ground truth has been collected; and there's the cloud....; but then again, personally only experience with unsupervised learning and smaller datasets, no idea how this scales up). Would be satisfied with 1-2 sentences though ^^.

Click to expand...

Ok, i give it a try. By deep learning we understand the approach to feed the computer trillions of games and let him analyse those sessions in order to find strategies.
Our approach is to use the existing knowledge humans have and combine it in a most efficient way rather than let the computer guess around often enough to find a suitable solution.
The most significant difference is that we think the creative ideas and inspiring solutions a human brings in are useful and should be used rather than be "washed out" by raw numbers.

Hope this helps more than it confuses

From my understanding, HIRO is an "Expert System" which does something like when Amazon says "People who bought this book also bought these books", only instead of basing it on customer data, an Expert System says "90% of people who encountered this situation did this action".

HIRO is probably more sophisticated than that.

Hans Lemurson said:

From my understanding, HIRO is an "Expert System" which does something like when Amazon says "People who bought this book also bought these books", only instead of basing it on customer data, an Expert System says "90% of people who encountered this situation did this action".

HIRO is probably more sophisticated than that.

Click to expand...

This would could only work if you knew every existing situation (like amazon knows every product). Think of a mouse set into a maze. First it tries out every possible way but with every try it learns more and more about the maze until it finds the shortest way. In this scenario there are only 4 options at max (go up, down, left and right) and they are applied at every tile of the maze. When to use what option is the experience.
Now we have a x-dimensional maze (called a civ map which is no maze) and hundreds of possible actions but the idea stays the same.

So we are talking neither of an expert system (it has more than just a weighted graph) nor of an deep learning algorithm (we dont analyse the games but just judge the success of the existing knowledge items)
Its hard to explain without getting too much in either the lecture or the marketing mode.

So the AI is going to get trained with the overall structures of decisionmaking required in a game of Civilization?

Hans Lemurson said:

So the AI is going to get trained with the overall structures of decisionmaking required in a game of Civilization?

Click to expand...

Yes, this is getting close to the real thing. A descision is a selection out of possible options. The human players give HIRO the possible options (like: a worrier who sees a worker should attack it). The Engine will learn to choose the best options over the time, getting a powerful player by this.