Print

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

[Reelya]

I didn't realize I wanted this until I started playing Aurora 4x, but I really wish we had a turn-based space 4x that actually followed at least a pop-science understanding of relativity. Even something as simple as speed-of-light lag on communication would be a dramatic difference from normal 4X titles, to say nothing of the absence of FTL.

As a practical matter, the galaxy would probably have all its objects move mostly on rails until the player perturbed them and then rely on patched conics, but even under those conditions, I think it would be fun to play in space as boring and lethal as reality.

I have thought about this on and off, since about the time Master of Orion 1 came out. The problem is that the concept is cool, however the game probably wouldn't be. If a turn is 1 year, and say that a probe can travel at 0.5c, then to send your first probe to Alpha Centauri would take 8 turns, you'd find out about what it found on turn 12, then you could send orders to advance to the next star system, which would take 4 turns to be received (turn 16), and another 4 turns to acknowledge the orders were received (turn 20). Then, it would take 8 turns for said probe to move onto the next star system (turn 24) and you'd hear about that on turn 32. So far, this isn't too big a problem. However, the real problems start to come when you are sending out ships that are supposed to be crewed by sentient beings. Unlike a probe, it makes no sense for a ship of humans to arrive somewhere, radio to base 50ly away and await a response for 100 years before deciding to do anything.

Other disconnects will creep in such as technology levels. Are outlying colonies going to have the old tech, and wait for new tech to be radioed-in from your home colony? What if the new colony ends up bigger than the capital, are they just going to put all R&D on pause while waiting for a distant start system to radio them with new tech updates? The problem is, this doesn't make sense either, so you need each group of sublight humans effectively doing their own R&D and ship redesigns, production, tactical and strategic planning, and not waiting around for "orders". The problem is that now, you must either have an entirely AI-driven "game" that almost entirely plays itself (which would be taking a more realistic approach), and it's not a 4X empire game after all, or you have a 4X god game with sublight communications where everyone except you is a dumbass who sits around twiddling their thumbs for decades while awaiting orders (which would be taking the straight forward "4x with sublight communications" approach).

Think: 4X is Star Trek or Star Wars, and if you add the sublight constraint, you've got Firefly or Cowboy BeBop instead. A different type of thing.

[Twinwolf]

I could see a 4x where speed is strictly sublight if it was limited to a single star system, maybe.

[Egan_BW]

You could do it, you'd just need a very robust AI and the acceptance that the player probably won't even see most of the things that happen, let alone have a hand in them.

Also, it doesn't really make sense to me that a drone would wait around doing nothing for 10 years. You know what situation you're putting it in, you can program it to get there, broadcast findings, then immediately move on without confirmation.

You could also go "best" of both worlds, with strict physics simulation within systems, and jump points or wormholes or whatever to get to other systems. No magical FTL communications, either. If you want to know what's going on past the wormhole you need to send something through and back. If you have a constant stream of data ferry drones cycling through the wormhole, you've got basically real-time information from the other side.

Ah, maybe my ideal space strategy game would be hard to do with turns. >.>
If the observation post you have next to the wormhole is one light minute from the neatest inhabited planet, I want that minute to matter. ALL intel is old intel. Some of it is just seconds old rather than years old.

[Truncatedurist]

Dwarf Fortress 1.0.

[Kagus]

Our small minds could not even hope to understand such a thing.

[PTTG??]

Using neural networks, make an AI that is extremely good at a turn-based strategy game, ideally to the point of being better than humans possibly could be.

Then, give the player the full ability to undo and redo actions.

[Reelya]

AI based on genetic algorithms evolving "rules" is a better fit here than neural networks.

NNs do pattern-matching and they do optimizing of function values. They're good for optimizing some continuous function or for dividing inputs up into categories. What they're not good at is coming up with high-level plans and sticking to them. And they're easily tricked via the process of creating "adversarial examples".

NNs can only get better at mapping inputs to "correct" outputs. If you don't know what correct output is, NNs can't learn anything. So a pure NN based game-player is only going to get good enough to beat the other really shitty AI players. They'll get stuck at in some local maximum of performance, such as optimizing the # of troops to send on a cavalry charge against spearmen, rather than thinking that maybe charging spearmen isn't the best strategy.

Instead of that, you can make an agent that's assigned a random set of "states" with random rules to transition between states, and random things the do in each state. Then set them loose and let survival of the fittest and mutation lead to less-useless agents over time. Eventually, they'll evolve e.g. a "seek food" state and have the appropriate triggers for when to enter that state, and seek the food properly. This process has nothing in common with neural networks, but it is machine learning, and is often part of the "higher level" controller stuff that might have neural networks embedded in it doing some low-level stuff.

[JimboM12]

firefly as a space game. has to have spacewestern music, both in and out of ship gameplay and lots of immersive nonsense to do like stop in bars and drink for buffs. crew management with both unique crew members with stories and things and mauve shirts random gens. ship modding with visible changes to the ship depending on the mod. people remark on how crappy your used ship looks but you know she'll fool em'.

[Niveras]

firefly as a space game. has to have spacewestern music, both in and out of ship gameplay and lots of immersive nonsense to do like stop in bars and drink for buffs. crew management with both unique crew members with stories and things and mauve shirts random gens. ship modding with visible changes to the ship depending on the mod. people remark on how crappy your used ship looks but you know she'll fool em'.

Red Dead Redemption in space?

[Trekkin]

You could do it, you'd just need a very robust AI and the acceptance that the player probably won't even see most of the things that happen, let alone have a hand in them.

Also, it doesn't really make sense to me that a drone would wait around doing nothing for 10 years. You know what situation you're putting it in, you can program it to get there, broadcast findings, then immediately move on without confirmation.

You could also go "best" of both worlds, with strict physics simulation within systems, and jump points or wormholes or whatever to get to other systems. No magical FTL communications, either. If you want to know what's going on past the wormhole you need to send something through and back. If you have a constant stream of data ferry drones cycling through the wormhole, you've got basically real-time information from the other side.

Ah, maybe my ideal space strategy game would be hard to do with turns. >.>
If the observation post you have next to the wormhole is one light minute from the neatest inhabited planet, I want that minute to matter. ALL intel is old intel. Some of it is just seconds old rather than years old.

Arguably if you can send matter through a wormhole you can also send light through it, so rather than a constant stream of drones you can just have a radio relay at each end, but traversable wormholes themselves produce another problem at the mechanical level: players are naturally going to want to move the termini at relativistic speeds even if the game rules enforce timelike separation of all paths through the wormhole network. Either the game pre-calculates the entire universe in a sort of inverse light cone, running the most distant regions out farther in future time, or we run the risk of someone launching a wormhole terminus into the void at high speed, poking their head through it, and looking at events that have not happened yet from the game engine's inherently privileged frame of reference.

Recalculating the future history of the universe every turn is going to get expensive.

[PTTG??]

AI based on genetic algorithms evolving "rules" is a better fit here than neural networks.

NNs do pattern-matching and they do optimizing of function values. They're good for optimizing some continuous function or for dividing inputs up into categories. What they're not good at is coming up with high-level plans and sticking to them. And they're easily tricked via the process of creating "adversarial examples".

NNs can only get better at mapping inputs to "correct" outputs. If you don't know what correct output is, NNs can't learn anything. So a pure NN based game-player is only going to get good enough to beat the other really shitty AI players. They'll get stuck at in some local maximum of performance, such as optimizing the # of troops to send on a cavalry charge against spearmen, rather than thinking that maybe charging spearmen isn't the best strategy.

Instead of that, you can make an agent that's assigned a random set of "states" with random rules to transition between states, and random things the do in each state. Then set them loose and let survival of the fittest and mutation lead to less-useless agents over time. Eventually, they'll evolve e.g. a "seek food" state and have the appropriate triggers for when to enter that state, and seek the food properly. This process has nothing in common with neural networks, but it is machine learning, and is often part of the "higher level" controller stuff that might have neural networks embedded in it doing some low-level stuff.

Take a look at what they've been doing with AlphaStar.

[Reelya]

There was some controversy over the results with that. It was pulling in data from the entire map at once, beating humans, but when limited to one screen at a time got its ass handed to it by the same players. AlphaStar could "see" the whole map at once and instantly be aware of any unit anywhere appearing from the fog of war. Additionally, it has a pretty high click rate, perfect accuracy of clicking and it never misclicks or makes mistakes. The real problem with AI for RTS is that it's hard to separate "clicking perfectly" to actually having a good strategy.

By all accounts AlphaStar has incredible micro, but it always produces units from a limited range of "normal" units. It's doing heavy number-crunching and lots of micro, but generally from a much more limited palette of options than a human would. This fits what I said about how an NN-based AI will "micro" the details on a cavalry charge vs spearmen but they almost never come up with a different "macro-level" plan.

Also read up about how AlphaStar works. Almost all the infrastructure that makes the system work is not part of the neural network. for example:

To encourage diversity in the league, each agent has its own learning objective: for example, which competitors should this agent aim to beat, and any additional internal motivations that bias how the agent plays. One agent may have an objective to beat one specific competitor, while another agent may have to beat a whole distribution of competitors, but do so by building more of a particular game unit. These learning objectives are adapted during training.

None of this stuff is internal to the neural network itself, it's all boilerplate code / genetic algorithm framework. By itself, a pure NN approach doesn't really do anything: they're not learning to e.g. make more tanks, some human decided that it would be good to have one of the AIs make more tanks, so they put "make more tanks" in as a constraint and merely killed off / mutated AIs that didn't fill their tank-making quota. Some human had to come up with the need for a particular AI that was more prone to making tanks however.

The approach taken in this case also suffers because their main learning is from the AIs playing each other. The big problem is that they can't learn from new experiences like a person can. They're only waiting to see specific triggers (NNs are pattern-matchers) and having hard-wired responses to those triggers, and the only feedback is whether you won or lost a match. this is where the genetic algorithm comes in.

[Kagus]

Except that the eleventh match also incorporated a lot of AI abuse, and the fact that it was a "fresh" AI and not even from the same roster as the ones that beat the players in the earlier games, in addition to testrunning the window view limitation, however it was they implemented that.

And, hilariously, the eleventh match was also the only one where the pro player ended up using one of the AI's strategies, which it had developed on its own and used fairly consistently.


Yes, there are issues with limiting machine resources appropriately (I mean, "omniscience" isn't an unreasonable descriptor for the map presence of some pro players), but it's unfair to say that AlphaStar only won because of "dumb fast clicks".

It is, however, always popular to take a diametrically opposed counterpoint whenever something makes headlines.

[Egan_BW]

Resident Evil Roguelike.
ASCII graphics, procedurally generated mansion, zombies. Manage resources and find keys opening up more of the mansion. Eventually you can go all the way down to the secret lab, activate the self destruct, and escape. While being chased by a tyrant, obviously.
Random enemies and weapons keep you on your toes, even a random assortment of bosses which can pop up at unpredictable times.
And if you waste too much time, a persistent enemy starts following you around, like mr. X.

[itisnotlogical]

I wish there were something like Daggerfall Unity, but for TES1: Arena. Make it less buggy and more user-friendly while changing very little of the actual gameplay. Arena is laughably primitive compared to Daggerfall, but I still prefer it (while loving both of course.)