|
This is for programmers/logic people, but ideas of others are welcome too.
Short Version:
Suppose u have a model of a starcraft player with skills like micro/macro/etc. (whatever u can think of) with values ranging from 0 to 99.
Im looking for a way to compute the winner of a starcraft match based on those values. (not the same player winning all the time, but that the better players wins a higher percentage)
Long Version:
The idea is to make a StarcraftManager browser game where u manage your team, train your players, make the lineups and have your teams compete in leagues.
matches between players will have to be calculated based on the various factors involved i.e. player skills, race matchup, map, used strategy (allin cheese, fastexpo etc), match situation (being down 2-0 in a BO5, ace match etc).
bonus:
it would be nice to have the generation of certain events that can be used to have a description of a played match.
example:
for the football browser game at www.hattrick.org the match engine roughly goes like this:
the playmaking and stamina skill of your midfield players make up your midfield value. compared to your opponents midfield value this determines ball posession. based on ballposession there will be "scoring attempt"-events for each team. based on the forwards scoring ability and the defender+keepers defending ability the success of a scoring attempt is computed (semi-randomly).
this will generate a series of reportable attempts and goals and finally a result of the match 
things you could consider:
player:
micro, macro, multitasking, timing, pressure handling, buildorders, experience, map knowledge, fatigue from playing/training
map:
number of starting locations, rushing distance, number of gas expoes
match:
pressure situation, previous games of these players
desired properties:
- a high skilled player always crushes a newbie
- a PvZ expert crushes a ZvT expert
- 2 equally skilled players, player1 goes FE, player2 goes allin rush, early in the match player2 has an advantage to end the game, players1 needs micro/pressure handling to defend. if he defends player2 is at a huge disadvantage
basic idea:
modelling the economy and the army (size, position, special units like DTs) and having them influence one another on regular events
or do you think this would be way too complex to implement?
|
I think there are way too many external factors to be considered that would make any system of this sort not as accurate as you may hope. Progamers get injured, get sick, etc., how do you account for that? How much of a handicap do you give to an amateur in a pressure situation? There is no way to give a concrete value to these things, and thus impossible to measure from 0 to 99. Your ratings will most likely be subjective in EVERY category, and you have to have data from ALL games played by that player for it to be accurate. I don't think you have the time or objectivity to collect data with that in mind. Also, the best player does NOT always win, you will never be able to make a model that will make that happen or predict the result. Sorry.
One solution I can come up with is something like this, although the entire concept of a model of this sort is weak in general.
You go back to the beginning of every progamer's (or at least the two you are focusing on in a match) career, and start them both at 0 points in all categories.
Every televised win is +5 points to player rating, +1 to pressure game clutch
Every televised loss is -5 points to PR, -1 to clutch
Every win vs. a person with a higher ELO is +7 points, +2 clutch
Every loss vs, a person with a lower ELO is -7 points, -2 clutch
Every game in a Ro8 has the points multiplied by 1.5
Ro4 multiplied by 2
Finals multiplied by 3
Every win/loss vs. race X is +/-5 points to vX
Each consecutive loss vs. player X costs more clutch (-1.5 for the second loss, -2 for the third loss, -3 for fourth, etc.)
Breaking a streak of losses vs. player X gives the amount of clutch you would have lost (+3 for a win after 3 losses)
Breaking a streak of wins vs. player X by losing to him takes away from you the amount of clutch he would have lost (-3 for a loss after 3 wins)
Consecutive wins would work under a similar principle
Etc. etc.
There are tons of conditions you can come up with, and I'm sure there is a way to do this in spreadsheet if you have little programming experience. However, this will still not be perfect and will not be able to predict the winner with anywhere near 100% certainty. If there was a unifying hierarchy like this (ELO doesn't really predict anything, Bisu vs. by.hero, etc.) then devoted people would be getting 99% on Liquibet because any chump could find this magical rock-paper-scissors model.
Also note how the system I mentioned would be taking every game, past and future, into account. Any model that does not do the same will fail. The other problem is how would it account for rises and falls of player skills? Kespa rating is nefarious for this.
Basically, this is very difficult and the results will not be what you want, but if you want something fun and not entirely concrete, go ahead and make it [:
|
|
|
Oh that is the kind of thing he was talking about?
-_-
FML
|
I think you're overthinking this, depending on what you're actually trying to do.
For instance, 1 and 2 of your "desired properties" can modeled simply by comparing ELO and competence in the matchup (i.e., 80% ZvZ in jaedong's case, etc). There is always a small chance a newbie will win, so this should be a linear scale normalised to 1 to determine the probability of victory. In my opinion, you should also consider map strength for a particular matchup. That should cover 90% of games, and all the data is available in the TLPD.
I suspect you're trying to win the fantasy proleague. In which case, I'll tell you how that's done.
What you're after is called a Monte Carlo (MC) simulation. It's based on a Markov model (or Markov Chains, or whatever you want to call it), which is just a fancy word for random process. Basically, you make a table of every possible matchup (every player vs every player, on each map), and record the probability of victory for each player based on that matchup. That gives you the Markov model. Obviously, the more complex you make the model (adding factors as you mention them above), the more accurate it will likely become. But what I mentioned is probably enough.
**You should be wary of people in TLPD with 100% matchup percentages because they haven't played many games.**
Dealing with small sample sizes is left as an exercise to the reader.
At this point, you can run a simulation. Assuming random matchups from each team in proleague (not true, obviously, I'll get to that in a sec), pick four players from each team and simulate the outcome of each game by randomly selecting the winner weighted on the probability of victory in that matchup (which you've already calculated). Run the match as usual; if the match ends up 2-2 have an ace match, etc.
From there, simulate the entire season, and you should have your team winner, as well as statistics collected on each player.
Do this several million times (shouldn't take more than a minute or two, unless you've done something wrong in your algorithm), and average the outcomes. Over time, you should have a very accurate idea of how each player will perform. Simply pick the "perfect team" from these data.
This is *exactly* how betting parlours calculate bets. There are incredibly complex models which are simulated many millions of times. It works, otherwise they wouldn't make money 
So, back to the matches. Obviously coaches aren't stupid; you'll have to assign probabilities that different players will be sent out on different maps. Another exercise for the reader.
And no, this isn't that complex to implement. I use monte carlo simulations for tons of things. The hardest thing by far is to generate the model. Once that's done, the actual simulation part is just a long for-loop with some decisions and logging functions.
EDIT: fanatacist brings up a great point: past match history between two players. Just add it to my small list of factors above
|
3 Lions
United States3705 Posts
You should look at the browsergame wc3manger.com...which has closed down and still hasn't returned my money >.<
Anyways, here's what I think the different stats should be.
Game Sense
Speed
Micro
Macro
Morale
Stamina
Experience
edit: are you a programmer? If so PM me and maybe we could try to make SCManager, as I have a bunch od ideas but don't know shit about programming
|
thanks for your replies
On March 26 2009 21:48 miseiler wrote:
I suspect you're trying to win the fantasy proleague. In which case, I'll tell you how that's done.
no im not doing this for fantasy proleague and also i dont want to win liquibet
also i dont want to calculate the winner of 2 actual progamers based on their historical stats.
the idea is to create a fictional browser game.... this is not based on actual progamers!
im just looking for a way to simulate a starcraft match based on skills of the players, the map, the match etc.
you can decide which skills players have and how much they influence the match engine!!
also it would be great to generate some sort of report for a match, so users of the browsergame can kinda review what was going on in the match.
On March 26 2009 21:50 3 Lions wrote:
edit: are you a programmer? If so PM me and maybe we could try to make SCManager, as I have a bunch od ideas but don't know shit about programming
yes im a programmer but this blog post is supposed to be the discussion platform
|
i wonder if anyone commenting here has ever played a football manager game. They have like 20 stats for each player. That's the feeling the OP wants to get I guess, a real comprehensive, possibly overly comprehensive rinse down of all possible attributes. PS there is nothing more complicated about Starcraft players' skillsets than football players, and they have made several successful football manager games ¬¬. It's probably quite doable even without a big programming team.
|
This is very interesting! I've been thinking about something like this but I lack the dedication and time required..
I think the model you describe is a very good start, yes it will be difficult to implement but no I dont think it will be too complex.
Some quick ideas/suggestions:
On March 26 2009 21:22 r3dox wrote:
player:
micro, macro, multitasking, timing, pressure handling, buildorders, experience, map knowledge, fatigue from playing/training
You should perhaps include scouting in there. Maybe something like:
In case of a rush:
If a scouting (should factor experience as well?) check fails, apply a certain multiplier (<1) to the players micro skills when defining if he will survive the attack.
Use the result to continue the simulation with the players' economies modified accordingly.
map:
number of starting locations, rushing distance, number of gas expoes
Rush distance should have various values ie for cross or close positions in 4 player maps etc
Also, it may get dangerously alike to MyStarcraft but race balance stats for each map may be difficult to avoid.
Honestly, the way a mach should be simulated depends on the level of realism and complexity you want the match engine to have.
You could go for anything ranging from:
Mix all the stats in a big calculation with the appropriate factors and randomness and display the result.
to
-Determine the initial bo of each player,
-include harassment events and battles using the micro, macro and multitasking stats,
-have the players expand depending on their advantage, risk taking etc
-do 2 again
etc
The danger is that the closer you get to the last method the more difficult it will be to get realistic results.
I also agree with fanatacist on the fact that it may get very subjective rating each players skills, especially for less known ones. The only way around this i can think of is to have a relatively large number of people who follow the scene (should be too hard lol) rate each player as objectively as possible and get the stats from that.
I hope this helps, I'm sorry for the text wall 
EDIT: lol i just noticed. Hi r3dox this is soup :D
|
|
|
Way too complex to model. It wouldn't be a very good predictor.
For the hack of it.
In this model with multipliers a glaring weakness is a heavy punishment. Every variable get a different weight for each stage of the game.
micro x (early game)weight x macro x (early game) weight x multitasking (early game) weight x timing x (early game) weight x pressure handling x (early game) weight x buildorders x (early game) weight x experience x (early game) weight x map knowledge x (early game) weight x fatigue from playing/training x (early game) weight etc. etc. etc.=Value
Devide Value player A with Value player B. Use a probabiltyformula to for a spread like a bell curve. Games won/lost by A in early game. Probabilty game hasn't won by either players.
A percentage of the games hasn't been won in this stage.
Repeat the proces with different weights for mid game.
You get something like
Early game:
Chance player A wins %+Chance player B wins%+neither%=100%
Mid game
Chance player A wins% x (% neither won early game)
I m not going to finish this because THIS IS MADNESS
But you can see where this is going at.
This is all I could come up with on the fly and I haven't done this for a decade or so.
|
Don't really get it, is it suppose to be be like fantasy football? fantasy proleague ROFL
|
|
|
|
|
|
EPT
