Metagame Analysis - Competitive Strategy
Toastie
Join Date: 2012-11-05 Member: 167351Members
Join Date: 2012-11-05 Member: 167351Members
I am working on a project to analyze NS2 from a machine learning perspective, with the goal of being able to predict outcomes of games based on strategic decisions. Ideally I would be using some form of clustering to discover strategies, but before I get to that step I need a better grasp on what the strategies are in the game and to develop algorithms to work with them.
I want to start with working with competitive data, as it has as it seems it will be much more consistent and understood than pub play. One of the big challenges is breaking down the data into strategies, so I would like some competitive community help with this. What I am looking for is more along the lines of small pieces (such as fast phase), than whole game strategies. Also the strategies don't necessarily need to be effective, just to be present in the metagame.
As a note, strategies which are easy to pull out of the data (such as fast phase = (armory, obs, phase tech) as first 3 in build order) are preferable to general one (early pressure = ???), but I will do my best to try to incorporate those as well.
Thanks for reading and helping.
<b>TL;DR:</b> what strategies exist in the competitive metagame?
I want to start with working with competitive data, as it has as it seems it will be much more consistent and understood than pub play. One of the big challenges is breaking down the data into strategies, so I would like some competitive community help with this. What I am looking for is more along the lines of small pieces (such as fast phase), than whole game strategies. Also the strategies don't necessarily need to be effective, just to be present in the metagame.
As a note, strategies which are easy to pull out of the data (such as fast phase = (armory, obs, phase tech) as first 3 in build order) are preferable to general one (early pressure = ???), but I will do my best to try to incorporate those as well.
Thanks for reading and helping.
<b>TL;DR:</b> what strategies exist in the competitive metagame?
Comments
In competitive play, strategy choices don't really counter other strategy choices. It's more about how the teams execute said strategy or the other team responds by positioning differently. It's nearly impossible to tell how this is going by looking at a build order. It's more about the actual tactics employed within a strategy instead of the choice of carapace vs celerity, hive vs carapace, mines and shotgun vs mines and phasegate, etc.
I suppose you can aggregate win/loss records given certain tech paths. I don't know how statistically significant the results will be given the influence of individual skill difference, tactical execution, and team skill difference which are somewhat difficult to quantify.
How do you plan on objectively evaluation how teams execute tactics and what a given response is? It's very easy to see what is built, but hard to quantify "pressure" and to differentiate that between other tactics. How do you plan on doing this across hundreds of games in a reasonable timeframe?
I'm curious to see where you go with this project. Please post updates if you continue with it!
- Team K/D
- Ratio of total resources gathered (i.e. Marine total res/Alien total res)
I suspect there are specific threshold values where the probably of winning approaches 100% (i.e. if the total resources ratio is 1.5 at 5min, then marines are likely to win), but I haven't had the time or access to enough data to determine them.
Basically what others are saying here is that the strategic layer in NS2 is not the primary determinant of game outcomes. Which kind of makes sense given the disparity in skill levels between players: if say player x has a consistent 2:1 k/d against other players, he is effectively a twice as powerful unit as the average player, which trumps say a 10% boost to damage or armour.
So what I think might be interesting to look it is analyzing game outcomes based on <i>statistical </i>analysis, kind of a moneyball approach to NS2. Eg graphing alien and marine win% agaisnt k/d ratio. Looking at win% based on resources generated, average # of RT's. Win% based on first fade timing, etc etc.
The strategic dimension could still be interesting: eg alien win% based on first evolution, statistically most successful marine win% tech path, etc.
Obviously ns2stats.org could be very useful source for such info. "Moneyball" of NS2 could potentially yield some very interesting insights, certainly for competitive teams.
I have access to the data, it's just not reasonable to try to work with at the moment.
Yes I realize player skill / KDR / total res / % map controlled are better classifiers for NS2 than strategies, and in the long term I will incorporate those as well. The goal of this is just to get above 50% (better than random) from strategies, and personally I am most interested in discovering information about strategies rather than simply predicting a winner.
those can easily map to nice feature sets for ML purposes
this will give you some part of the 'strategies'
i.e. when were mines/phasegate/arms lab researched, when was 2nd hive dropped, when was 2nd cc dropped, which hive evolution did aliens pick and when
there are really only a handful of viable combinations (these may change / be wrong... just illustrating the point)
like marines is going to be some ordering of mines / shotguns / phase tech / arms lab / 2nd cc
aliens is going to be some ordering of hive upgrade / 2nd hive / tres lerk
anything beyond that is probably just noise because the game is usually decided by then
- Team K/D
- Ratio of total resources gathered (i.e. Marine total res/Alien total res)<!--QuoteEnd--></div><!--QuoteEEnd-->
No, this isn't really true.
Now, you might argue that the variable BO is an operational definition of the variable strategy, but if you operationilize strategy like that, you miss out on a lot.
Let me give you an example:
Some time ago, both archaea and exertus played veil using the same opening BO. We both went fast PG, but we used them entirely different. Ex used them to get a foothold on the map, and then they used it to get alot of RTs up and then they just turtled until fully they had all tech they needed. Turtle strat.
When I used that PG, I used it to do a timing push on the 2nd hive. Timing push strat.
Same BO, completely different strategies. In a game like this, in general I think "qualitative methods" is preferable to use over "quantitative methods" here.
That is the reason why he never drops medpacks!
Given you got enough data you can probably distinguish between player skill and strategy, thus answering usual questions like: "how much k/d explain win/loss?".
P.S. Onos > Hadoop elephant.
Res lost/gained
RT towers lost/gained
Hives lost
Lifeforms lost
Guns lost
K/D -> Depends on the engagements. 1 marine dying to 3 skulks should hurt the win ration graph, however 3v3 fight should do.
Upgrades timings (set to defaults with threshold to compare against)
Decision making on upgrades (again, compared to some defaults. E.g. when the lerk pops up - whats better? sg->dmg2? dmg2->sg? Agree on default and compare with what marines chosen)
Alien upgrade loss.
Some other stuff...
So the win ratio is 50/50 at the start.
When marines build 2 RTS on each side, and aliens get 1RT out it will become 55/45, as the rt ration is 2.5:1. When the second alien RT pops up, the win ration should change to 45/55 respectfully. When you loose a marine/alien -> ration changes. Team fights have to affect win ration much higher than single engagements vs single/multiple targets.
When you go for dmg1 upg. and aliens have celerity -> marines should not gain much of the ratio++, as the fast dmg1 is good against cara, so in that case SG choice would be better? or armo1? (you have to collect a lot of meta game data, before having those CONSTANT default strategies). However, countering carapace with timed dmg1 upgrade should affect win ration. The same upgrade but timed not properly (5min instead of 3min) should not gain the same win ratio+ for marines.
It is not THAT easy to do, as you want to make it - a lot of AI coding lol. But it is achievable I must say. I know how AI coding works.
Proper rules and facts can be achieved, but as I said - initial proper analysis for those facts have to be done.
I mean, casters or any viewer can predict outcome of the game by engagements, upgrades, etc, etc. Why AI shouldn't?
Will definitely look forward for that mod :) Good luck.
This is a motivation of the work in part, and thanks to the NS2stats team, I should have access to plenty of data. The computing power to deal with it on the other hand...
Anyways, I've finally gotten around to getting the data into objects, started writing the simple methods.
At this point I have:
KDR / spread
Tres gathered
Next I'm going to do the build order, RTs built /lost, and try to look at some Pres spending and losses (that could be a bit tricky, I'll see what I can do).
Unfortunately I am still a ways off of doing a lot of the map-specific stuff, as that requires me to translate the geometry of the map into something I can work with, but I will definitely get to that. I'm also considering how to identify a turtle or push to see if I can factor that in as well.
Also if you can you should also keep K and D as variables.
I would be curious to see a KDR vs win/loss scatter plot.