So... can you 'beat' this game?

[Texashawk (OP)]

What I mean is, does anybody (or has anybody) played this game to its end, for good or bad? Steve's AARs seem to end basically when he... gets bored or creates a new version, whichever happens first. I love this game, and I love the AAR and novel elements and I am aware that this system was originally created as a medium to tell stories, but has anyone actually conquered the galaxy? Is it even possible? Is there a way to win (or a victory screen) that I'm not aware of, or is this sort of the Dwarf Fortress where you decide what winning (and losing) looks like to you.

Just curious what all of your experiences with 'winning' Aurora have been.

-Steve

[Erik L]

Winning is when you go long enough to conquer everything before the database crashes

[SteelChicken]

Not really.  Getting bored is the usual method of ending a game.  The NPR's just never provide a challenge.   I think there is something wrong with the NPR generation...20 years in a game I went into to designer mode and all the NPR's were worse off than when they started.  5-7 Research labs, still at TL1 or maybe 2.  All out of money and minerals. 

[Thundercraft]

Getting bored is the usual method of ending a game.  The NPR's just never provide a challenge.   I think there is something wrong with the NPR generation...20 years in a game I went into to designer mode and all the NPR's were worse off than when they started.  5-7 Research labs, still at TL1 or maybe 2.  All out of money and minerals.

Are you talking about your experience with v6.21? I'd like to hear whether or not other players are experiencing the same thing. If so, this suggests that there's something wrong with the AI control of NPR's. That can be a serious issue (especially for single-player or solo games) as it would cause the game to rapidly get boring. Having very weak AI opponents has ruined the reputation of more than one 4x game and strategy game. In such cases, if it were my game, fixing the AI to an acceptable degree of challenge would be my top priority.

On the other hand:
Aurora does have the Difficulty Modifier adjustment on the new game configuration page. It's 100% by default. If games aren't providing enough of a challenge, you might try increasing that and see if that helps. That is supposed to give the NPR's a larger starting population, a faster research rate, and a faster population growth.

That said, I usually take issue with games that cheat too much in order to provide a challenge to players. IMO, a good AI should be capable of at least providing a decent challenge to new players without having to cheat, such as automatically knowing where all the jump points are, where the best planets are to colonize, etc, without having to invest some time or resources to search for such. If that is necessary to provide a challenge to experienced players, so be it. But the AI is very often a weak point in games.

If I'm loosing a game to an AI with a level playing field, that's fine. But loosing to an AI that had an enormous head leap or that cheats like a 2nd grader will just make me feel like... I was cheated.

[Erik L]

Are you talking about your experience in version 6.21? I'd really like to hear whether or not other players are experiencing the same thing. If so, this suggests that there's something very wrong with the AI control of NPR's. That's a serious issue (especially for single-player or solo games) as it would cause the game to rapidly get boring. Having very weak AI opponents has ruined the reputation of more than one 4x game and strategy game. In such cases, if it were my game, fixing the AI to an acceptable degree of challenge would be my top priority.

I think, if nothing else, giving all NPRs major advantages (like extra research labs and periodically injecting more minerals) could be a temporary fix.

PS: Having AI opponents that cheat too much and are more of a challenge that most players are comfortable with is also a potentially big issue. Though, having options to adjust game difficulty greatly improves both problems.

The AI is pretty limited in Aurora. I do know it is on Steve's to do list.

[Paul M]

The problem with an AI is that for the most part they are scripts. 

They lack the ability to analyse a situation and the certainly can not question the information they have.  This means they don't ask "why is that section of the line weak?"  They just determine it is the weak spot and attack.  In a game like HoI3 that means the AI falls blindly into encirclements a child would notice.

They lack the ability to learn from play.  This means the only time the AI will surprise you is the first time you play or the first time you encounter a specific situation.  After that you will know what the AI is doing and adapt to it.  Since you can learn and adapt and the script stays constant the challenge it can provide drops dramatically.  Using HoIx again, that is why it was possible (generally also by gamey exploits) for Tibet to take over the world using only militia.

They lack the ability to plan and manage.  This is critical in a 4X game.  Strategic planning takes a few seconds for a human:  you decide you are going to "use beams and big ships" for example.  After that though comes the real stuff where you go into management mode to queue up the things you need to implement your strategy.  RTS games or for that mater most "strategy" games, aren't about strategy...this is hold over from chess books talking about chess strategies, which aren't strategies.  RTS games or just about any game that calls itself a strategy game are about management.  "The science of using men and matrials efficiently" to quote the canadian armed forces.  But AI's in these games can not pick a strategy, they follow a script.  They cannot adjust their strategy, they follow a script.  And worse they can't adopt their management of resources to fulfill their strategic vision, not that they have one in the first place...they follow a script.  For this reason the AI generally needs a different rule set than the player, as this kind of thing is pretty much impossible to program or script.

Also the problem for programing an AI is the question of "why?"  Why did you build unit A and not unit B?  Why did you attack here and not there?  Sometimes the answer is trivial but most of the time when you really ask yourself that question the answer is complex and it involves, for example: your current forces, your projected forces, your naval lift, an overall operational objective and your commitment to a different theatre.  You can program a chess playing computer program because chess is a very "linear" game with limited options and constrained board.  For the most part it is tactical, basically the further in the future you can see the more chance you have to win.  This is where the computer does better than a human, visuallizing the board arrangement 4 moves ahead for every possible move you can make is possible but it is hard. 

At the end of the day, making an AI that can provide a challenge after the first few plays to a human is very very difficult.

[Thundercraft]

The problem with an AI is that for the most part they are scripts.
...
They lack the ability to analyse a situation ... can not question the information they have.
...
They lack the ability to learn from play.
...
They lack the ability to plan and manage.
...
Also the problem ... is the question of "why?" ... For the most part it is tactical, basically the further in the future you can see the more chance you have to win.  This is where the computer does better than a human, visuallizing the board arrangement 4 moves ahead for every possible move...

Those are excellent points. And I think you pretty much nailed the problem on the head.

Thing is, artificial intelligence is a science. There are advanced college courses devoted to it. However, my impression is that many game designers - particularly indie developers - do not devote much time to the AI. Some don't even crack open a book on the subject, much less have a college course on the subject under their belt.

I'm just saying some developers seem to treat the AI as an afterthought, whereas it should be a major focus. Granted, you have to have a game under development before work on the AI can progress. But too many publishers prematurely release a game with dismal AI. And sometimes (not always), some time down the road, they release a patch to improve it.

As you said, a game of chess is very linear. But a game like Aurora is very complex, with a lot of possible "moves" or choices being analog or even abstract. And a lot can depend on the game options and settings. A strategy that works beautifully in some settings may tend towards failure in other situations.

How does one go about programming the AI? Creating a "script" for the AI to follow seems like the only viable solution. But that's the easy way out.

A somewhat better solution would be to have a number of different AI "scripts" and have each computer opponent randomly pick one. That alone would make it much more difficult for a human player to predict the moves of computer opponents. Also: In the Allow Limited Modding of Aurora thread I mentioned how fans of VGA Planets were given the ability to use pseudocode to write their own AI script files to replace the default AI.

Taking it to the next level, a framework could be developed for the program to literally learn to improve itself, perhaps relying on one or more scripts initially. A good AI can learn from AI losses and wins/gains, from player losses and wins/gains, or AI updates from a central website.

It's true, though: Electronic games have been around for, what, 30 going on 40 years? And in that time game AI hasn't developed a whole lot. What more modern games should be aiming for, especially with the memory and processing power we have today, is a basic simulation of a neural network to learn the way biological organisms learn.

I mean, come on! Computational models for neural networks have existed for OVER 70 YEARS! It's successfully utilized in speech recognition and image analysis software, and in autonomous robots. So why don't we see it utilized more in modern games?

At the end of the day, making an AI that can provide a challenge after the first few plays to a human is very very difficult.

You make it sound as if creating a decent AI for a game more complex than chess is so difficult it borders on the impossible.

While the quality of AI in games as a whole tends to lean towards poor or pathetic, there are exceptions. I'd like to consider Galactic Civilizations I an exception - it's actually too challenging for it's own good - but the computer opponents cheat so darned much (they all have an omniscient knowledge of the entire galaxy) it's difficult to tell how much of its ability is a simulation of strategy and how much is cheating. The reason I mention it is because the company boasted of how the AI "learns" by connecting with their website and literally updates itself with data from the games of other players, training itself to emulate the techniques of winning human players.

It's a bold idea and I wish more games did something like this, though I'm sure it'd be a challenge to program and the mileage would vary. That, and I have a hard time trusting programs that - behind the scenes and without prompting - connect to the Internet to share my information, regardless of the reason.

[3_14159]

The problem, as far as I see it, is not only the lack of AI development (no criticism here! It's long and tedious work, and the current model works for at least providing a short-term solution), but also a problem of game scope:

Good AI requires an environment designed for it.

Take, for example, the old FEAR game. There is an interview (which I cannot find at the moment), in which one of the designers specifically said that the AI appeared better than it was, for example using flanking tactics, because the levels had been built for that. While the AI could use these levels without special adaption, it only shone in ones that had special characteristics.

Now, for Aurora, the problem is that you have big, interconnected puzzle parts, some hierarchic, some not. For example, take an upper-level management AI trying to both research and build a new class of ship. You have to decide what priority should be given to researching the components of the ship. So, this upper AI may task a ship builder with building a specified ship class (having received a request for that from a military planning AI), and get a number of research projects back, which it must then chose to research.

Therefore, probably the best way would not be using one (or n) scripts per NPR, but several. Something like that:

Master AI: Sets priorities for the rest.
Foreign AI: Foreign threat assessments and possible responses
Military AI: Fleet Deployments
Fleet AI: One per fleet, specifies actual tactics during fights
Economic AI: Colonies, Buildings, Productions, Research. Possibly one per planet.

And all of them have interconnectivities (well, most), need to be programmed and debugged, and so on. It's a hell of a lot work, that's for sure ;-)

[Paul M]

Any human who can read this can beat a scripted AO "artificial opponent" because at the end of the day the script is limited.  The difficulty only comes when the AO is given the ability to cheat.  I spotted this first in MOO when I reduced each of the computer empires to a single planet.  Suddenly a warship and a colony ship showed up over them...and died since I had a garrison at each world.  But that explained how the computer was able to simultaneously expland and have a big fleet.

AI reserach is at the point where they can program a computer to play a game which is extremely formal.  There exists a codified set of openings and counters to those openings.  There exists codified ways to win with combinations of pieces.  There are clear heirarchies of moves.  Chess tactics are well developed.

The problem with games that incorporate strategic considerations is that changes the game for you.  Should you chose to not use missiles, then everything about your naval design changes.  In HoI what your research...what you build...all of this is an outcome of your few seonds to minutes of strategic planning.  It then is adapted to fit your means, and what happens.  You also constantly assess and re-evaluate the situation...and more to the point you make plans.  It may well be the invasion is moved up or the 3rd Fleet ends up in the Atlantic or your fuel farm flatlines but still you had at least some thing that was a plan.

All of these things are exceptionally hard to overcome short of having a real AI present.  The chain of IF-THEN-ELSES you work through automatically is very complex and often involves judgement rather than a strict rule adherance.  90% of the time you do A, but because of Factor Alpha, and condition 2 this time you do B.   Based on experience you adapt or change things.  I don't know if it is impossible to provide an AO that is challenging without cheating but I would suspect that for a 4X time game it is exceptionally difficult.

[alex_brunius]

The problem with an AI is that for the most part they are scripts. 

They lack the ability to analyse a situation and the certainly can not question the information they have.  This means they don't ask "why is that section of the line weak?"  They just determine it is the weak spot and attack.  In a game like HoI3 that means the AI falls blindly into encirclements a child would notice.

They lack the ability to learn from play.  This means the only time the AI will surprise you is the first time you play or the first time you encounter a specific situation.  After that you will know what the AI is doing and adapt to it.  Since you can learn and adapt and the script stays constant the challenge it can provide drops dramatically.  Using HoIx again, that is why it was possible (generally also by gamey exploits) for Tibet to take over the world using only militia.

They lack the ability to plan and manage.  This is critical in a 4X game.  Strategic planning takes a few seconds for a human:  you decide you are going to "use beams and big ships" for example.  After that though comes the real stuff where you go into management mode to queue up the things you need to implement your strategy.  RTS games or for that mater most "strategy" games, aren't about strategy...this is hold over from chess books talking about chess strategies, which aren't strategies.  RTS games or just about any game that calls itself a strategy game are about management.  "The science of using men and matrials efficiently" to quote the canadian armed forces.  But AI's in these games can not pick a strategy, they follow a script.  They cannot adjust their strategy, they follow a script.  And worse they can't adopt their management of resources to fulfill their strategic vision, not that they have one in the first place...they follow a script.  For this reason the AI generally needs a different rule set than the player, as this kind of thing is pretty much impossible to program or script.

Also the problem for programing an AI is the question of "why?"  Why did you build unit A and not unit B?  Why did you attack here and not there?  Sometimes the answer is trivial but most of the time when you really ask yourself that question the answer is complex and it involves, for example: your current forces, your projected forces, your naval lift, an overall operational objective and your commitment to a different theatre.  You can program a chess playing computer program because chess is a very "linear" game with limited options and constrained board.  For the most part it is tactical, basically the further in the future you can see the more chance you have to win.  This is where the computer does better than a human, visuallizing the board arrangement 4 moves ahead for every possible move you can make is possible but it is hard. 

At the end of the day, making an AI that can provide a challenge after the first few plays to a human is very very difficult.

Indeed.

What the developer can do however is add AI randomness.

As a developer you should ask, what AI behavior can a player easily detect and predict, and how can I add randomness in such a way that the player can never predict for example what weapons, strategies or attack point the AI will use.

This will often give the illusion of a thinking AI, even if it's just picking weapons, ranges and strategies at random. And it will always mean it's hard to counter the AI in other ways then brute force/technological advantage.

I think Aurora does this fairly well for NPRs, but Spoilers turn out to be very predictable once you learned how to defeat them.

In a realtime game the AI also has a major advantage over a player, it can act/react instantly everywhere at once. This advantage disappears in Aurora due to the game being turnbased with unlimited thinking/action time for the player. In Aurora this advantage translates into the AI having an easier time to manage a huge empire efficiently (if programmed the right way), as larger empires quickly will become hard (or prohibitively time consuming) to manage for a player.

[SteelChicken]

Are you talking about your experience with v6.21? I'd like to hear whether or not other players are experiencing the same thing. If so, this suggests that there's something wrong with the AI control of NPR's. That can be a serious issue (especially for single-player or solo games) as it would cause the game to rapidly get boring. Having very weak AI opponents has ruined the reputation of more than one 4x game and strategy game. In such cases, if it were my game, fixing the AI to an acceptable degree of challenge would be my top priority.

On the other hand:
Aurora does have the Difficulty Modifier adjustment on the new game configuration page. It's 100% by default. If games aren't providing enough of a challenge, you might try increasing that and see if that helps. That is supposed to give the NPR's a larger starting population, a faster research rate, and a faster population growth.

Yes, my most recent game I started with 200% difficulty.  I will be monitoring the NPR's in designer mode and make sure they dont run out of minerals or wealth.  The last game before my most recent I had them at 125%, but they were still 2-3TL's behind me by the time I made contact and that was with me giving them bonus minerals and wealth on the back side.

I started playing aurora with version 5.x, and I remember it being more of a challenge, but then again I was a newbie player.

[Conscript Gary]

Even without higher-level decision-making ability, AI reactionary decisions could be scripted fairly successfully I think. Take missile design. I'm not sure how Steve has them programmed, but I suspect they use constrained random designs according to your own tech. As we all know, designing for yourself is a good start but what lets a player wipe out an AI decisively is the ability to design for your enemy. If the AI instead designed its missiles with known alien ships or missiles as a target goal it would increase their bite and reaction ability significantly.
The trick with such a thing is that it doesn't have to depend on AI work. Just taking the mathematical analyses people have made on the forum, refining them a bit, fuzzing them randomly so that not every missile is the optimal design, and there you have it.

I make a hell of a lot of assumptions in this post as to how things work, but I stand by my opinion that poring over the actual numbers and equations behind the mechanics is critical from a balance standpoint and can be useful to emulate decision-making.


I want to further articulate my point with an example, a relatively simple one that should be familiar to many of us.
Scenario: We've spotted a new ship, and it's father than anything we've seen before! Holy cow! Do we need to design new missiles?
As a human,  one might answer this simply. You would check whether it can outrun your standard ASM outright, and if the to-hit chance is still acceptable. If it's still good, you carry on. If it's deficient, and you want to be able to hit said ship, you design a new one. There's numerous judgement calls there, but for a human it's manageable.
As an AI, the first stepup is similar enough. 'Can it outrun my missiles' is a simple comparison of speeds. But how do you decide what an 'acceptable' to-hit chance is? That depends on how many missiles you carry, how many you volley and how fast, and plain old simple taste. Given those factors as racial attributes, if we can make some weighted numerical metric than the AI can indeed answer that question.
Let's say our current ASM falls below that calculated tthreshold, so it's time to fire up the missile design code. You have your target speed and target to-hit percentage. You have your existing missile engines, your standard launcher size, standard engagement range, and you churn all of those cconstraints and parameters to get the closest you can get to that optimal design. Is it impossible to attain? Check backwards to see if a new engine design would satisfy your parameters, and if that fails try while projecting other tech advances or constraint changes. Weigh those results with their cost (In time, RP, resources, doctrine momentum) and go after the one that's best, even if that's 'no change'.

...It's a hell of a lot of factors to account for, and I even glazed over some. Properly reducing design decisions mathematically gets tricky for more abstract  aspects. Still, I also think that randomly-racially assigning those 'taste' decisions to a more extensive level will also enrich the game.

I think this topic of how to improve the AI might warrant its own thread sooner or later.