Battle Calc Question
-
@Cernel Valid concerns, however I'm convinced that if we code it the right way (so it essentially uses the same code as normal battles) the performance improvements will be huge at no cost at all.
Simply because the code currently simulates battles by copying complete savegames, running the battles there, storing the results and then undoing all the changes.
I'm sure this will reduce the overhead of TripleA drastically -
@roiex I still suggest 1 year time of keeping both before deciding if throwing the simulator overboard.
-
@Cernel There's no reason to keep 2 systems doing the same thing around for such a long time. Either the potential new system works as expected and we keep it and dump the current one, or it doesn't work as expected because stuff isn't working as expected and we dump it instead.
Simple as that.
Nobody wants to introduce a half-broken system just because it's lightweight -
Btw it's easy to forget we can in theory revert all changes that were ever done to the triplea codebase thanks to git.
We can even lookup what the first code being written 16 years ago was exactly -
@roiex shush you
-
I've thought a lot about about calculating actual probabilities, and it looks impossible with aa attacks and support, but I probably could find an approximation that would be fast and fairly accurate in battles with hundreds of units and no support/aa stuff (but inaccurate with fewer units). That doesn't look useful enough for me to try it, though.
The tree search looks computationally impossible if battles can last 4-5 turns (if there are only, say, 4 different dice outcomes in each round for each player, that's around 4^10 ~ a million nodes already).
On the other hand, a machine learning algorithm should perform quite well here. You would need to train it for each map separately though, which is an issue, because you need a lot of computational capacity to train a neural net. But once trained, it would be fast, reliable and accurate. (Actually, for maps with 10ish units, it might be pretty cheap computationally to train one?)
You could also combine the tree method with heuristics. You start building a tree, and if you reach a node where one player is just obviously going to win, or if a node is deep enough down the tree, or it has a very low probability, instead of recursively continuing the tree, you apply a simple heuristics based on hitpoints/power to evaluate the node directly.
-
@alkexr Interesting Ideas.
Will think about them and how feasible they are. -
@alkexr I don't think a neural network would need to be trained for individual map.
If we think about a unit as a tuple of an attack and defense value, cobine it with an amount count (simply being 0 if this kind of unit doesn't exist) and try to handle any special events with a classical aproach, the only thing we'd need to do is to limit the maximum die value maps can specify (might be very high if needed), and pass in the die amount of a map as an additional input. -
@roiex said in Battle Calc Question:
If we think about a unit as a tuple of an attack and defense value
Except how do you tell the net which support type affects which unit, and which AA attack can target which unit. And units can have multiple rolls. And multiple hitpoints. And we can have coastal bombardment, and capturable combat infrastructure. But if we ignore these, I guess that's possible.
No, I think you need to have two inputs per unit type: the number of units of that type on each side. And then you have to let the net learn how combinations of units present affect each other. (Except that triggers, technologies and terrain can change unit stats too. Erm... whatever.)
But for some simpler maps with a limited number of unit types this might be viable. You would only have several hundred neurons. Now I never actually tried training a net myself, but I don't think that should take longer than an hour on a PC. And it would massively improve performance if the map tends to have huge stacks of units. As for more complex maps... not sure.
-
@alkexr Yeah, I agree that unless you found a way to aggressively trim the tree that it would expand to quickly for battles that can last 7-8 rounds.
@alkexr @RoiEX I kind of doubt that without a lot of data and a lot of effort that machine learning would work here. There again are probably too many different parameters and the variation across them is pretty large. Honestly, I think simulation is the best approach and improving the simulation that we have today is the best use of our efforts.
-
@redrum said in Battle Calc Question:
I kind of doubt that without a lot of data and a lot of effort that machine learning would work here. There again are probably too many different parameters and the variation across them is pretty large. Honestly, I think simulation is the best approach and improving the simulation that we have today is the best use of our efforts.
This is actually a comparatively simple problem for machine learning. We already know the correct solution to the problem (battle calculator), we have an unlimited amount of training data (battle calculator), the number of parameters is relatively low (hundreds at most with the map specific approach, and that is for the most complex maps) (compare that to hundreds of thousands for image recognition), desired output is a single number (win chance) that changes more or less continously with input.
Although I agree we shouldn't focus too much effort here. Especially since it's probably not going to be me who is going to sit down and do some actual work
-
@alkexr You might want to try it though,we're always looking for new developers
Hello! It looks like you're interested in this conversation, but you don't have an account yet.
Getting fed up of having to scroll through the same posts each visit? When you register for an account, you'll always come back to exactly where you were before, and choose to be notified of new replies (either via email, or push notification). You'll also be able to save bookmarks and upvote posts to show your appreciation to other community members.
With your input, this post could be even better 💗
Register Login