We've just put up a video trailer for Three Kingdoms. Link

Originally posted by CustodianV131:

Just after CA hit with Total War Three Kingdoms. Is that good or unlucky timing? Also wonder or this would do better as a standalone.

We've known for a long time that the TW game was coming out, but people keep asking for 3K content for OE, so we're obliging, and really if you're doing a game about China it's something you've got to do sometime anyway, so we just decided to cheekily ride on their hype a bit. But anyway The Three Kingdoms is a great story and a great period for gaming, and it's great that CA are doing so much to popularize it in the West.

@The 4Xplorer. TW games are made with at least 20 times the budget of OE, not surprising that they can do something hugely difficult and time consuming like diplomacy a little better.

Seems to be lots of irrelevant reasons why the first car wasn't as good as a horse.

The goal of the project is to develop new AI techniques, not to develop a world champion starcraft AI. They've demonstrated an AI that can play the game to a very high level, a level that's far higher than is necessary for any commercial game AI.

The fact that it runs on an "outdated" version of the game is completely irrelevant. Presumably this is because thats the version of the game they have that has a convenient AI interface. If it were being built for a commercial game then it would play the version of the game that it shipped with, the same as all existing game AIs. And since the training period is only a couple of weeks, when game rules are tweaked the AI can be updated too.

Similarly, the objection that it only plays one race is trivial to overcome. Fire up a training regime for each race, create a set of agents for each one, and load as appropriate when the game starts.

Cheating is similarly irrelevant. All game AIs have access to the complete state of the game. The difficult part is putting that knowledge to use. A useful game AI doesn't have work on completely level playing field with the player, it only has to not obviously cheat.

It would be interesting to know how important the single map restriction is. Obviously a random map adds hugely to the state space of the game. There's no obvious intrinsic reason that an AI shouldn't be able to learn to explore and identify key points, but I'll agree that they haven't so far demonstrated this. However, even if that's a major stumbling block, many complex strategy games are played on fixed maps, or a small set of maps.

There are two things that they've done that make me excited (and surprised) about this. Firstly they've demonstrated that's possible to train a neural network to play a game as complex as Starcraft to a high level. Impressive as the exploits of Alpha Go were, Go is a game with perfect information and a few hundred possibilities at each turn. Until this news hit, I'd assumed we were still a very long way from being able to tackle the huge state spaces of computer strategy games.

Secondly, they've demonstrated that computer hardware has advanced to the stage where it's possible to do large scale competitive multi-agent training in a reasonable amount of time. No doubt the kit used for training is expensive, but it's not beyond the budget of the larger studios, and hardware costs are falling all the time. It's also possible to rent pretty much as much computing power as you want from Amazon and Google, including GPUs.

There's obviously work for game developers to do to adopt these techniques. The NN structures used need to be studied, training objectives for agents will similarly to be researched, and game engines need to be rewritten to allow high performance AI agent play. But if I worked for Firaxis or CA I'd really start looking at this seriously.

The linked article is rather lacking in detail HERE[deepmind.com] is the article on the Deep Mind website.

In particular it makes two points. Firstly that the AI did not use APM that were higher than the humans it played against, and in fact it's average was lower. And secondly that subsequent to the matches they have already developed a version that has the same view constraints as a human and that it's almost as strong as the unconstrained one.

In any case, it's irrelenant whether or not there currently exists an AI that can currently beat the top Korean players. What Deep Mind have done is just to reveal a giant signpost in twenty foot high glowing neon letters as to how game AI will be developed in the future.

Pointing out ways that it still can't match humans is rather like pointing out that the first car wasn't as fast as a horse.

This is an incredible achievement and a deal quantum leap in how to go about creating AIs for complex games. Their paramaterization for Starcraft has ten to the power of twenty six possible moves each turn, whereas Go has a few hundred. Whether or not Starcraft is the most complex game out there is irrelevant, the leap in capability is amazing, such that it's only a matter of time before it's possible to have self training AI's for even more complex AIs where APM isn't a factor at all.

Basically it points to a (fairly near) future where instead of being painstakingly programmed by hand, AI's will essentially be grown in a computer farm where they learn by continually playing each other. It also appears that the technique they use of having multiple agents with own agendas could give quite good control over the resulting AI such that you can create agents with a desired playing strength.

The kit they use for training will be a bit expensive for most Devs who aren't Firaxis, but I guess you can rent sufficient computing power from Amazon.

Of course the core AI algoritms should be part of the base game, but AI isn't some well defined system that you build once and never touch again. Good AI requires a lot of detail work, tweaking and testing.it's an excellent candidate for post launch improvement. During development the gameplay and balancing is continually evolving, there are a limited number of testers, and nobody is yet sure what the best strategies are. Once the game gets into the hands of thousands of players, the amount of knowledge about how it plays, and about how the AI copes, explodes exponentially. This new knowledge could be tremendously useful in informing further AI development, but realistically most post launch development effort is going to be channeled into features that players are willing to pay for.

Ridiculously ambitious. If you want to compete with Elite or Star Citizen then you'd better have a budget well north of $10 million.

HF is certailnly heavy and full of science, but it's also a great game if you have time and patience. What I particularly like about HF is the sense of achievement you get when you pull off a mission. Random events can slap you down pretty hard in your first few games, but usually with hindsight you see that were ways you could have avoided them.

I'm not sure there's any definite plans for the fourth edition, but the authors do mention it from time to time on the Yahoo group, and probably BGG too.

Surprised no-ones mentioned The High Frontier. Fantastic hard sci space colonization game. Pretty much 4X although a bit light on extermination. Downside is that it's out of print, but they're planning a 4th edition next year probably.

Originally posted by DevildogFF:

Originally posted by Imirak:

Anyway, I’ve invested way too much energy in to Dan. I don’t think I’ve believed in an indie game as much as I did StarDrive 1, and I can honestly credit it (and ES1) with the reigniting of my passion for 4X. However, Dan lacks honor and is very egocentric. I will no longer devote any of my time or energy towards him or his ideas.

Moving on!

Though not of course without buying Longsword just so that you can leave the only negative review with 0.1 hours played.