quote:Of course it is. Any computer program is limited to a single responce under a specific set of circumstances*
Well, I don't think that the AI should necessarily be limited to one course of action.
Yeah, but that means different courses of action, each of which is determined by circumstances. The circumstances can be reduced to a bit pattern representing game state, and any decision to pursue a course of action is a function of that bit pattern. Combinations/sequences of actions included.
It's perfectly fine to at least discourage micromanagement, so as to keep a level playing field between player and enemy AI. That's part of the spirit of the game.
Each person looks for different things. Personally I find the above statement in direct conflict with what I am looking for in this game. To me it reads "since the opponent AI is dumb, let's force the player to endure the same dumb AI on his own units". It is much better to achieve play balance by giving the computer opponent numerical or sitiational advantages (like time limits on the player), than to force him endure watching his units blunder into destruction.
As it's unrealistic to assume the devs will produce a HAL-like AI, there has to be an achievable approach. If you want to micromanage to the extent where you maintain an advantage over enemy AI, sure, that's your call. I'd prefer the playing field to be as level as possible, while still allowing as much scope as possible for player 'command level' decisions. I don't like the idea of AI 'cheating' as so many games do, thus the decision trees have to be comparable to what a human can do. My perception is that CMANO aspires to minimise micromanagement by automating such behaviour as much as possible. The challenge is to make that automation as realistic as possible.
Yep, these are our subjective opinions on what the game should be, but player feedback's good, right?
quote:I would say impossible to implement. Computer AI only survives the scrutiny of human logic when hidden behind fog of war. Once fully exposed to the player's view no computer AI will ever be able to behave in a way that is not frustratig to the player, at least as long as we don't have a whole different computer technology. CMANO, like Harpoon before it, avoids the worst part of computer AI deficiency by breaking it down to relatively simple missions and putting the burden of giving some semblance of realistic reactions on the scenario designer.
To keep that level playing field, such actions must also be available to the enemy AI, as well as an acceptable judgement/decision logic that can decide which to perform in what circumstances. I reckon that last part is by far the trickiest to implement.
I agree, but still a useful level of decision processing could be achieved if enough factors are taken into account. I reckon that's part of the challenge of finessing the AI - adding more factors to the decision process so as to produce a more sophisticated result. It will never achieve human-level performance, but clever design means it shouldn't have to, to result in a challenging game. Good scenario design will be important.
quote:As an example of how quickly this breaks down is tha fact that the definition of "close to enemies" can be interpreted a hundrded different ways. Does the enemy have maritime patrol aircraft? Helos? Then immediately the definition of "close" as a matter of distance changes. Add to that the fact that while the computer has a strict model of uncertainly as to the position of enemy units, the player has a much wider and very subjective way of determining where he thinks enemy units are. So it is effectively impossible to program the AI to "run on diesels until close to enemies" because the AI has no way of determining what the player considers "close to the enemy" and it is therefore doomed to not act the way the player wishes.
For example, here, one action 'profile' would be to run on batteries until a charge threshold...
Sure, I only gave those examples as they've already been discussed, without going into any detail of how they'd be implemented. Any 'profile' would consist of many behaviour components, each of which would be determined by game state, preference settings and decision logic. That's what I mean by hammering out what we want our units to do when acting automatically, so as to decide what would those components be, and what would determine their enactment.
BTW I did suggest a means of judging being "close to the enemy" here.
quote:All that is a huge amount of programmig work and even bigger amout of troubleshooting and beta testing. And it can all be simply avoided by a checkbox in the speed and altitude screen. Scenario designers can be given the same control over AI units so the can truely adapt the computer AI tot he scenario circumstances in a way generic AI never will.
However, how would the enemy AI decide between A OR B?...
I agree it could be a lot of work - that's why narrowing it down to a useful subset is important. Then the relevant player preferences could be implemented, including via checkboxes. If the AI is to compete realistically, then its decision trees should be able to choose the same preferences by some means, including by scenario design.
quote:That is impossible to 'hammer out' as the different circumstances would take pages upon pages and include things like "how much time do I have left", "how close am I to winning the scenario", "what are the scenario objectives" as well as pure player preference in tactics.
Maybe we need to hammer out just what behaviours we DO want, and be precise about WHEN we want them, so as to work out the decisions required if the AI is to mirror them.
I still think we could reduce it to a useful subset, though that would mean rejecting factors we judge to be least important. And yes, that is subjective, but that's where the idea of player feedback would matter. Ultimately, though, it's the devs' call, and their guidance would be crucial.
Any improvement over the current 'single course of action regardless' would be a plus.
quote:In what world would the problem of "The AI is acting dumb" be solved by "instead of the AI acting dumb in a consistent way let's make it act dumb in a random way".
If any part comes down to pure player preference instead of decision-based logic, then that part could be a random factor for the AI.
* Unless you make the AI randomply pick from a set of possible responses, but that is a horrible idea for player side AI (great for computer AI though).
That's not what I mean. If certain player decisions cannot be reduced to determinism/rationalism, then they're effectively random, even if weighted. "I'll send this patrol over the sea, because I like the colour blue." If the computer does the same, then it's no different. Dumb? Maybe, but that's humans for ya. In any case, our world is regarded as non-deterministic, if the scientific consensus on quantum mechanics is to be accepted. Probabilistic decisions could be useful and realistic.
FWIW I think this particular sub issue would make a good test case for trying to achieve a better automated result. If that could be achieved here, then so too for any other area where the automation could be improved. But until that happens, manual override might be the best short-term fix.
< Message edited by guanotwozero -- 2/1/2014 6:32:07 PM >