It was mentioned that the (now so-called) AI will reach its maximum skill at the "Normal" level. Above "Normal," advantages are given to computer players so they can play "better." (All quotes are mine.)
This is very disappointing. I had higher hopes for this game.
Perhaps it's because AI is my career, and I believe no game should need to cheat to win (or at least tie). I think the team decided to give up before the job was finished. I don't mean to be harsh.
Maybe I'm alone on this. I'm genuinely interested to know what other people think, no, feel about this situation. Some people care that the physics is modeled correctly, others that black holes look realistic. I care about the...hmm, not sure what to call it...although the term AI is overused today, I hesitate to use it when it applies to what might just be a deterministic set of if...then...else rules, or whatever is being used. It's not like I expected anything like "learning" behavior, but I did expect heuristics and other techniques and algorithms in sufficient quantity and quality to produce the same playing field that real world systems thrive in every day. Why is the bar set low for a game like this?
Is anyone else disappointed?
I read Gilmoy's last post whereupon my brain's primary circuit breaker tripped, my head tilted to the left, and spittle fell to the floor. In the nanosecond prior to the overload, I was certain I understood everything about the universe. Alas, all the gained knowledge exited my mind and body via drool and when I woke up, I was as dumb as before.
As far as atoms the current estimate thst i know of is one times ten to the ninetieth power of atoms in the universe.
As far as credentials i grew up on battlrstar galactica. Sorry to hear about lenard nimoy.
People make it way more complicated than it is. Ten trillion cells in the human body that are copies of each other. There are only 150,000 genes with chemical placements. Okay the chess programs aren't intelligent, and i scaled it down for the game.Let's begin with processing. You store a limited number variables which have to translate to a on of off switch somewhere. Stupid rlectronics! To make things you are limited with programming which only works in basically three forms first perform a set of instructions and exit, the second is a loop untill either done or false, and the third is to test untill false decision.
Has anyone pays attention to little kids? How comethey communicate very simple. Basically i reasoned the same at four years old as iI'm reasoning hear. I just had to start with step one.
Any took english what, where, when, why, and how. What is this that is coming in. I search my data base. If i don't find it instead of locking up label it the closest thing with one exception, and store it in a different spot. If it is identical instead of restoring it just add one to how many. Where did i find it, or where is it from. When did i find it, or when is this important. Why do i know this, or why is this important. How did this information happen. If there are different ways to analyse nyhis add it. Make each piece of detail as simple as possible. If different store different. Can this be rearrange d differently to form knew ideas. Is this unhelpful or helpful. You would need two copies of every event. One would he similar to a movie for description purposes. The other would be a database for the rest. How the information is imputed would affect this. Also how the information is disectedbfor the database is important. The more detailed database the better.
Remember it learns by whatbit experiences. Even this is a machine iteventually reacts by what it experiences, and told by repetitions. A value is given a string. Instead of refering to the value it uses the string. The program only knows the string not the value. would have no.way to access the value. A good example is yellow. If people usr a range instead of specific value then you would only know it by range. This would be based on repetitiveness of memories. Preferences and friendships would develope. please disect this logic. You could create a time limit where if not.relearned you will loose parts of the event in entervals that is acessedbwhen it is sleeping, but why. This would work for a computer, not a game.
Now comparing this to the show caprica just downloading all the records on the person is not enough. How the information is processed is signifigant. You might need to hardwire some things like it is alive, and no one can convince you otherwise. a fear of being alone. Emotions should develope naturally, but if they don't just put a value on certain kind of event's. You could hardwire right from wrong. You could make it situational for sbput four years, and then give it reasoning.
Now applying this to a game avatar instead of a machine. It already knows things like tactics. It would mine data as small as possible. Each avatar does this seperatly. It would mine the same tactics from each other, and you. It would compare the new tactics against its own. You already suggested two copies of the game. Storing the superior tactic. This could be reset at the start of each game, or even better remembered. If remembered the avatar wouldn't be able to hold the previous game against you so much it would use tactics to prepare against this in case you do. Even testing random stuff or copieing your moves. To keep this from being an exploit. Require advanced espionage, diplomacy, or visible confermation. If lose use your tactics by data mining or randomly doing something new competing against it's own tactics, untill it works. Giving a reset button in options for when you want to start over.
Just remember things are redundant like weapon values. Values on buildings not specific buildings. Speed and range values. Planets can be classified by class, and how you utilise random resources for observational purposes. Resources are random, so the conputer wouldn't need to know type of resources. As far as starbase buildings it could compare type of planet economy production research level approval and population of planet you build it by. Class of planets you skip or colonize. Stats of ships you build and what you do with them. Classify gameplay events. Types of technology tesearched. Types of buildings you build verses approval and population of your planets you build. What you do with starports, and you sponser them. Instead of comparing spaces on a board working with redundant variables instead. Comparing the recorded aspects ofbyour empire against it's own, and try to reproduce the infrostructure.
...
It would compare the new tactics against its own.
That is the problem as i understand. How does someone, or an AI does compare without giving values to the points to be decided?
The wrong assignment is where all decisions, human or artificial fail. Also values change fluently and the only saving grace humans have are experiance (could be simulated ... maybe, i doubt that) and intuition.
I learned the hard way to compose a long post in a text document. My sympathy. Of course sometimes a short post gets long apparently all by itself, it seems.
Well, I know the feeling.
What I get from this is normal AI= Challenging AI from before.
Don't see the problem, outside of maybe confusing vets of the series who aren't paying attention to this, and think Challenging is the fair level.
I really do wish Stardock made AI that would adapt most of the time to what a player likes to go for.
Well to be certain stardock has defined tactics. Planets can be defined by type, class, random events, resources, and what's built on the planet. How much population. Assuming advanced espionage. What can be built in starport.
Your right you need numbers. Factors on a planet can be devided into military, social production, research, and economics. There are numbers on each one per planet. There can be a range for non specialization. If one percent is teached for specialization?
Add the types of planets together, and create a percent for each one. Count the number of planets colonized. Planets are devided by type and class. If players skip over planets that are a certain type and class. Ships usually come in types that are classigied by mods.
Count stats. Empires have overall stats. There is a pattern to research. The game already classifies tech types. number and average.
Pay attention to race does the numbers change with different types of races. When you war what ships do you build, and what kind of planets do you attack . Include population in this. How do you fleet up. If this is what your talking about. What kind of ships player's go after.
If random things appear do player's react, and how. This usually defined by if they send or build ships. Add this for each game. Add total devide each section number into total. Their you have percent of what for different classes that stardock have a better list for. Compare your totals versus other player's.
About the other the machine needs a number the programdoesn't. You are dealing with a program. If string equals a certain range of values use strings. Unless i tell it what i'm doing it doesn't know. Ask a dtengin how old it is. Does it tell you how many many weeks it is. Even if it knew this which it doesn't unless i add the parameter to thr program i could still give a preference by either ehat it was taught first, or how many times people explained how to do this. Remember ypu are dealing with a program. If told value equal this then do this. If never told then your right. This is a way to fool you. Remember all i have to convince you Im intelligent.
So you have a bunch of numbers. But how do you even start to weight them? Which numbers are important for the one decission at hand. Should i choose a suboptimal path, because it would further my long term goal?
I think my point is: we as humans dont even understand how our own brain decides. We are able to map where decissions are made, what parts of our brain are included in the process. But we still dont know the how and why. As long as this is the case, the blind (we humans) are trying to teach someone (AI) the meaning of colors.
All we can do is to tell the AI how we would have decided in an special case and then find as many cases where that decission might be also applicable. But no one case will be exactly as the one before in a game with so many variables, so decission failure is pre-programmed (ha!).
I think Stardock does a very good job with his AIs, but i got beat up by the Normal AI in GalCiv2 sometimes, so maybe take my opinion with a grain of salt.
I think the greatest weakness of the AI in GC2 was that it didn't hold anything in reserve. You could stomp its planets and it would react in a very limited way.
A human player though, when faced with such an invasion, fights tenaciously. They can marshal their forces and send them where they are needed in strength, instead of diluting that strength with a number of limited-strength strikes against strategically-unimportant locations. When they go looking for a place to strike back they make sure that they know what lies ahead of them so there's less chance of flying into an ambush.
I'm not sure that you can exactly program tenacity into an AI, but you can make it a bit less interested in sending lone ships to be blown up one at a time. From what I remember, the AI in GC2 would certainly send ships to rally points to fleet up, but the rally points tended to be too close to the action, resulting in the aforementioned 'single ships go boom' scenario. When I built up a strong manufacturing capability I would send ships to a rally point roughly equidistant from the planets making ships, fleet them up and then send them to the front lines.
Well, in GC3 there are space shipyards instead of planet-based starports, and they can crank out ships with the combined manufacturing input of their sponsor planets. I don't know if that improves matters, but it's a step in the right direction.
I’m sorry that you are disappointed, let me clarify things a bit and see if it makes you feel better.
It may have been misleading when I said the AI will get bonuses at higher difficulties. While this is true, it is not as simple as +100% to shields credits per turn or anything like that. In fact the cruder bonuses of that type will really only be used in Godlike, where we want even very advanced players to feel like they are threatened.
The primary way the AI will be differentiated by difficulty, is by something we call AI Strategies. AI strategies are had crafted, by us to make sure the AI is playing to win. Each difficulty will have its own set of AI strategies
On normal, these strategies will help the AI play a rational, realistic game, they will not necessarily be out to crush you right away, and will in general be easier to deal with. If you wrong them they will try their best to kill you, without any extra help from us.
Beginner AI strategies will make the AI a bit less aggressive, and easier to push around by experienced players. We may in fact give them a bit of a -% bonus to production if needed, because we don’t really want to make the AI stupid, timid is fine, but stupid is unacceptable.
As you go up in difficulty, the primarily difference will be the strategies, they will become more aggressive and more clever. And by aggressive I do not mean, just motivated to win, in their preferred way. As an example the Krynn will try harder for an influence victory, or the Drengin will try harder for a Conquest victory. This is done by make more advanced AI Strategies. These are kind a blueprint for victory that the AI works with. They are created by us, so we can update them as players get more advanced, or figure out strategies that we have not accounted for.
At the highest level, Godlike, we will give some bonuses, but we are trying to keep them simple, things like making the AI more efficient builders, or slightly more accurate. But once again the real difference will be the AI strategies. I hope allays your concerns a bit. We are very proud of how the AI is advancing and we do not want you to think we are just toughing bonuses at the AI player, we have done a lot of work to make sure these strategies are flexible, and smart, and can be updated as the game matures.
Also, sorry I am so late to this thread, been kind of swamped around here.
Hey, thanks for checking in, it is very encouraging to hear from you on this topic. I just want to say to be careful with the bonuses. My biggest frustration with FE:LH is the dodge bonus that is given to enemy troops at higher difficulties. Some bonuses are really unfun to play against, and I end up playing that game with the AI at levels that aren't very challenging for me because I don't want to deal with that one bonus. Giving the AI advantages is fine, just try to be sure they are advantages that are fun to play against and overcome.
Please let us know what level is AI uses all algorithms but gets no bonuses?
Is that still Challenging?
Quoting mormegil, reply 85I hope allays your concerns a bit. We are very proud of how the AI is advancing and we do not want you to think we are just toughing bonuses at the AI player, we have done a lot of work to make sure these strategies are flexible, and smart, and can be updated as the game matures.
I like what you are saying and it is more than I expected.
My experience with AI vs human is the lack of dynamic. The AI can be programmed to win with a relentless attack, but the nature of AI is status quo. When the AI goes to war it attacks with the goal of taking all your planets before you take all of theirs. They are okay with losing all their planets as long as they take your last planet before you take theirs. The AI never stops to think that attrition is taking too many resources and victory may leave it too weak to continue against other opponents. The AI can be coded to sue for peace, but normally only if the opponent is stronger.
The goal of the human player is to take all of the AI planets without losing any of their own,
The AI is relentless and static.The human is dominant and dynamic
The experienced human player gets bonuses too in how it fits ships.
The experienced human player gets bonuses in how it tweaks planets.
Why shouldn't the AI? Why shouldn't it be abstracted as such? Just so someone can say, that AI is doing something the 'real' way, whatever that means in a computer game. As long as it's still competitive and challenging who cares. It's when it no longer becomes competitive and predictable there is a problem, all the bonuses in the world won't make it less predictable however.
The only possible way for an AI to outplay the playerbase just using behavior long term, is for people to constantly update the AI forever, literally forever with an unlimited amount of permutations. This is also why true conscious AI will be forever impossible, because infinity cannot be modeled well. Which is what it would take to accurately simulate a true conscious AI. No matter what you pre program, even conditions for learning, a human being, certainly collectively communicating ones, will adapt to. They will either come at the problem from a completely unexpected angle making your program obsolete, or find something you've missed, in life this is a hundred times worse as life changes also, but at least in a game the variables are somewhat fixed.
For the purposes of the game, giving the playerbase the tools to tweak the AI would help, as at least that'll be humans trying to outplay humans. I am not sure how feasible that would be though, but it'd certainly be helpful for modders. As long as people understand that the only reason those modders made the AI better was, player feedback and extended development / adaptation past the lifecycle of the game.
I would add multiple AI's are what usually make games, for me, interesting, as they throw up unexpected situations, and the more you add, the more complex the puzzle becomes.
There are many great features available to you once you register, including:
Sign in or Create Account
