Having a series of triggers is too much coding? I disagree entirely, HOI3 already does this and many mods do, its just on a small scale and not focused long term.
The Ideal AI
- Thread starter GermanPower
- Start date
-
We have updated our Community Code of Conduct. Please read through the new rules for the forum that are an integral part of Paradox Interactive’s User Agreement.
I am not disagreeng with your points here, nor am I trying to debate real AI. We are talking about programming an automated opponent, the term AI is only used for convenience. I still hold that doing so through a series of pattern recognition loops gives much more flexibility than the current system and that it is also not harder to code (unless the current AI is using 90% borrowed code that already exists).
Yes. I think we are in agreement. Extensive events and scripts should be drafted to account for as many ahistorical events as possible. It is theoretically possible. Having made a mod, I have learned a thing or two about what is possible with the Hoi3 game engine. There is a plethora of triggers, effects, scopes and modifiers built into the engine that are never used in the base game. Events are used sparingly and just to keep the game on historical rails. The event chain is relatively easy to break and if broken the game does not know what to do. If Italy conquers Austria before the Anschluss, there will be no WW2. Other bits of scripting is hidden elsewhere. Germany always observes the phony war and will wait until April of 1940 to attack towards France. If on the other hand, if Paradox were to code event chains and scripts for every "what if" they can think of and every exploit taken advantage of in every AAR, then the game would be on rails, but the rails would be so varying and complicated that no two games would end up the same and it would appear as if the game were thinking and reacting. Of course this would mean a 2016 release date and it wouldn't really work that well until after the third or fourth expansion.
.. different type of timing .......... :rofl:
Real time for a CPU register is a 3rd of a billionth of a second.
Real time for the lvl 1 cache is around 1 or 2 billionths of a second.
Real time for memory is around 10 billionths of a second for sequential data, and about 100 billionths for random data.
Real time for me and you is about 0.2 seconds
Chess, HOI and monopoly are ALL insanely slow for a computer.
Even though it looks simple to you, solving the chess board is extremely complex, not the way you play it but the way the AI plays it. Looking massive numbers of turns into the future. These cause huge numbers of branched paths (far more than units in an average HOI game). These are all calculated "simultaneously" from a humans point of view.
The HOI AI controls a larger number of initial units than chess but preforms far far less processing on each.
From a computers point of view they are similar.
Both have a small number of pieces in play HOI (~10000) chess (32).
Both have incredibly slow cycle times of around a second, compared to a billionth of a second.
These days Kurzweil is more a technological evangelist than a scientist/engineer. If your interested check out Andrew NG and Geoffrey Hinton, and google also have a superstar network guy whos name escapes me (but he will be linked to those 2).
I think your talking about a form of neural net, and your right it would be awesome for HOI. It is how our brains work.
But it would be a very difficult job. I would guess 2 magnitudes more difficult than the current AI.
Also Paradox would need to buy/rent a few supercomputers to train the neural nets. And even with that the trained NN's would hit HOI's performance significantly.
Hopefully in HOI 9.
I knew you would use that argument, so let me elaborate further. I hope you´ve played both Company of Heroes and Starcraft as it makes things a bit easier to understand.
In the baneling versus marine example you posted, the thought process is essentially this:
Move/avoid attack move/avoid attack move/approach attack.
It´s MANY simple operations, done a LOT of times. No wonder the human fails hard - how could he choose 20 marines and give individual orders to each? Impossible - but the CPU can. The human HAS to Split it into smaller groups, and even then fails hard. Truth be said, ever since the 70s (maybe even earlier) computers can do many simple operations at a same time at a ratio MUCH higher than a human.
Now compare that to COH:
1- Unit do random damage instead of rigid DPS - random factor
2- Units do diferent damage on the move - fire on the move factor
3- units do diferent damage at different ranges - range factor
4- units can use cover - cover factor (naturally also implied hat you also have to make a decision between staying in cover or moving if your DPS is higher at close range)
I´ll stop right here because it´s already darn obvious- the more variables you add to a move the harder time the AI will have. In other words - COH is a chaotic system in a much more complex way that SC 2 EVER is. So you shouldn´t compare doing na AI for HOI using Starcraft 2 - but COH,which is much more closer to it in the fact that is has MUCH more complex processes happening at the SAME time.
Instead of DPS, speed and range.
In the baneling versus marine example you posted, the thought process is essentially this:
Move/avoid attack move/avoid attack move/approach attack.
It´s MANY simple operations, done a LOT of times. No wonder the human fails hard - how could he choose 20 marines and give individual orders to each? Impossible - but the CPU can. The human HAS to Split it into smaller groups, and even then fails hard. Truth be said, ever since the 70s (maybe even earlier) computers can do many simple operations at a same time at a ratio MUCH higher than a human.
Now compare that to COH:
1- Unit do random damage instead of rigid DPS - random factor
2- Units do diferent damage on the move - fire on the move factor
3- units do diferent damage at different ranges - range factor
4- units can use cover - cover factor (naturally also implied hat you also have to make a decision between staying in cover or moving if your DPS is higher at close range)
I´ll stop right here because it´s already darn obvious- the more variables you add to a move the harder time the AI will have. In other words - COH is a chaotic system in a much more complex way that SC 2 EVER is. So you shouldn´t compare doing na AI for HOI using Starcraft 2 - but COH,which is much more closer to it in the fact that is has MUCH more complex processes happening at the SAME time.
Instead of DPS, speed and range.
There is one thing that most chess AIs have in common with HOI. That is scripting. There is so much established material at grandmaster level for chess openings. Most chess programs incuding the one run on Deep Blue to defeat the world champion essentially have these openings scripted. They have memorized the books. It is only once the human takes the game off the rails by deviating from established lines (including transpositions) that a chess AI really starts thinking for itself.
That is one of the tricks a human can use to take the lower end chess programs, that we can play on PCs, out of their comfort zone. You make a weak and meaningless first move like a3. The computer does not have this in its opening library and is forced to go adlib right away. This is similar to playing Romania and derailing WWII by attacking or even just declaring on the Czechs.
One of the problems with the events in HOI in the past is that many of them are setup in chains. This is why I still think a very simple web (or net if you prefer Concept's term) is preferable. Such would look for situations and react rather than simply making dice rolls against events which are considered valid by what has been defined as their trigger conditions.
I know they are not going to completely reinvent the wheel here. My suggestion is hypothetical and not realistic. They are used to the way they have done things and they will almost certainly be looking to refine what they know rather than starting over.
I think you are overestimating the complexity of what would be needed. We don't need an HOI opponent with whom we can converse on the intricasies of Plato or Chartre. It doesn't need to learn everything on its own and can be told a lot of it based on the programmers' knowledge.
Granted, your version would be an opponent that would learn to play better over time. I'd settle for his idiot cousin who at least plays competently most of the time.
It is the difference between a simplistic neural net and an adaptive neural net. I don't expect to see the latter applied to games in my lifetime, but with the accelerated pace of things I can't even be certain of that.
That is one of the tricks a human can use to take the lower end chess programs, that we can play on PCs, out of their comfort zone. You make a weak and meaningless first move like a3. The computer does not have this in its opening library and is forced to go adlib right away. This is similar to playing Romania and derailing WWII by attacking or even just declaring on the Czechs.
One of the problems with the events in HOI in the past is that many of them are setup in chains. This is why I still think a very simple web (or net if you prefer Concept's term) is preferable. Such would look for situations and react rather than simply making dice rolls against events which are considered valid by what has been defined as their trigger conditions.
I know they are not going to completely reinvent the wheel here. My suggestion is hypothetical and not realistic. They are used to the way they have done things and they will almost certainly be looking to refine what they know rather than starting over.
I think you are overestimating the complexity of what would be needed. We don't need an HOI opponent with whom we can converse on the intricasies of Plato or Chartre. It doesn't need to learn everything on its own and can be told a lot of it based on the programmers' knowledge.
Granted, your version would be an opponent that would learn to play better over time. I'd settle for his idiot cousin who at least plays competently most of the time.
It is the difference between a simplistic neural net and an adaptive neural net. I don't expect to see the latter applied to games in my lifetime, but with the accelerated pace of things I can't even be certain of that.
Last edited:
That is probably what we will see as far as events are concerned. The aspects of AI concerning things like production will probably follow a preset plan with some alternate plans that might be triggered under certain conditions. These preset plans will vary by nation. This is essentially what they have done in the past.
As far as events we have been given a couple of teasers. They have said that there will be fewer events triggered by dates and more that are triggered by certain conditions being met.
I think the bigger issues are the ones that are not trigger driven. The AI has to be able to make competent plans of what to do with its forces, especially air and naval. If we see big improvement in those areas it should turn out to be a good game.
Well it may be scripted AI, but then improve the bloody scripts. Sending 3 ships on the Channel to be shredded by the French or bombers over England when it has its fighter force intact are the classical examples of retarded moves that shouldn´t be tOO hard to avoid.
Also, HOI 4 should probably use AI switches like HOI 2 did, in a way or another. Or at least let us mod them and make AI follow them.
Also, HOI 4 should probably use AI switches like HOI 2 did, in a way or another. Or at least let us mod them and make AI follow them.
haven't really modded AI in either HoI2 or 3 but isn't the LUA AI script in HoI3 much more versatile than the switches from HoI2?
The answer is simple, HOI 3 had german ships on the channel or near Scapa Flow in september 1939, HOI 2 didn´t.
I'd like to see a mix of V2 and EU4 - AI sets itself a goal (EU4) and then tries to fulfill conditions to "activate" it (V2), without additional constraints. It should allow for flexibility while not making choices made by AI completely senseless or based purely on a dice roll. Add to this AI that acts better on tactical level - as Beaga says, tactical AI left much to be desired - and I'll find it a vast improvement over what we had in HoI3.
It might be more like 80% but they are reusing major parts of the logic. Heck I still saw HOI1 logic in EvW when I worked on that project.
That is why PI makes a big issue about redoing things like strategic air forces. That is completely new code. What I'm not sure about is how the Battle Plans will impact all of this.
But to get more specific I think we should discuss either strategic (who do I attack and when, what do I build, research etc.) vs. what I call tactical which to me is the movement of land forces and the actual battles.
Pattern recognition is not as important in the tactical area. The logic here is more along the lines of straight computation of odds and damage.
Now here are some other things that greatly impact "AI". Combat is done on an hourly basis and most of this is formula based. At midnight each day the "AI" gets an update on the 'state' of all units production etc. So once per day the AI sort of plans out what it wants to do. Each hour a much more limited set of instructions are carried out for possible battles etc. Finally, each month there are additional things checked and tested for. All of this was done for performance reasons. The logic is actually stored in two places. The LUA files and other parts are hardcoded in the EXE. This is how the engine works and even the latest Clausewitz engine operates the same way.
The Germans in the real war did the same as HOI3. The problem is how naval combat is done and not the fact that 3 ships sortied.
It IS part of the problem not using the enemy force to acess WHEN to send ships or planes. UK fleet was overwhelming in 1939 with the french fleet.
So no, it IS stupid to send the Kriegsmarine away. That spotting and other things don´t work very well, ok, I agree with that, but don´t say there isn´t an AI issue when it´s OBVIOUS there is one.
And if you don´t agree then show me games where human players dare send anything outside Germany in 1939. If raiders worked, was because they were small, single ships, not the 6 ship or above fleets the AI loves to send.
So no, it IS stupid to send the Kriegsmarine away. That spotting and other things don´t work very well, ok, I agree with that, but don´t say there isn´t an AI issue when it´s OBVIOUS there is one.
And if you don´t agree then show me games where human players dare send anything outside Germany in 1939. If raiders worked, was because they were small, single ships, not the 6 ship or above fleets the AI loves to send.
Last edited:
What I'm saying is HOI3 didn't work, we agree on that. But it didn't work because of many things and not just because the AI sent 3 ships out in 1939 or 1940.
After having spent some time on the air power DD I'm getting optimistic that HOI4 will solve this problem.
After having spent some time on the air power DD I'm getting optimistic that HOI4 will solve this problem.