|
Post by krebby on Jan 19, 2018 14:41:57 GMT -5
I've read your initial post about four times I still can't understand the following key thing:
What, exactly, is your "model?"
I get that you're trying to create a model that predicts the outcome of a match based on various things, including tankers and campers. But what is the model?
Is the basic model that: (a) given the same number of players; (b) equal equipment for each player; (c) equal skill for each player; and (d) perfect map starting point balance—will result in a 50% winning chance for each team?
Thus, using, e.g., a camping set-up or inappropriately using a range set-up would generally indicate lower skill?
|
|
|
Post by ᎶƦ℮℮ƊᎽ ƤΛƝƊΛ on Jan 19, 2018 14:48:37 GMT -5
I've read your initial post about four times I still can't understand the following key thing: What, exactly, is your "model?" I get that you're trying to create a model that predicts the outcome of a match based on various things, including tankers and campers. But what is the model? Is the basic model that: (a) given the same number of players; (b) equal equipment for each player; (c) equal skill for each player; and (d) perfect map starting point balance—will result in a 50% winning chance for each team? Thus, using, e.g., a camping set-up or inappropriately using a range set-up would generally indicate lower skill? It isn't indicating lower skill, it's indicative of less total bots being available for that team. And the model determines outcomes by essentially flipping a coin to eliminate bots. The premise of the entire model is that every single bot available is in the playing field at the same time. Based on that, I don't feel the model can be relied upon. Something that factors in the fact that 6 bots from each team are in play simultaneously, and neither team is at a disadvantage until a player(not any number of bots) is eliminated. Then if you want to use it in relation to tankers, you have to somehow discover the efficiency of a tanker as compared to your average player and have a variable for that. It's much more complex to determine by building a model, because there are many factors that need to be considered, some of which are entirely unknown.
|
|
|
Post by ᒪΛᏟIΛ on Jan 19, 2018 15:44:06 GMT -5
Yes for the cost of my 10 gold
|
|
|
Post by voyager on Jan 19, 2018 18:03:28 GMT -5
Tankers cause win rates of 20%. That isn't an assumption or model. That's data from thousands of actual games from actual tankers. So even if campers work out to 29%, it is a difference between that and 20%, not 25%. I'm still not sure where that number comes from, as I can't get any proof of it or see any data that backs it up. Dr Yat provided no data, simply said he had data. I've never seen your data. The camper thing would also better be figured out from an actual player camping in one bot intentionally. I understand the premise of what you're getting that number from, but the starting point is this magical 25% that a model was put together to obtain the predetermined number. You normally build a model and see what comes out and adjust variables from there. In this case, you picked an outcome, and made your model spit that out. The application of the model represents what would happen if all 30 bots faced off against all 25 bots at the same exact time, and that 100% of games are determined by which team mechs out first. But as far as I can tell, using your assumption of campers causing 29% win rates, and tankers causing 20% win rates(based on actual data from tankers), and assuming a balanced matchmaker forces a win rate of 50%, we get the following: Campers cause a 42% deviation from the normal win rate. Tankers cause a 60% deviation from the normal win rate. Based on this, we can conclude that tankers are approximately 43% more harmful to a team's chance to win than are campers. I know we've discussed on Discord but I feel it would be interesting to open up other opinions from the forum. Panda, forget the tankers for now. look at the logic for "yat" to draw that conclusion. its wrong. he played 20 games and won 10 and lost 10. that means campers have absolutely no affect on the outcome. Yat proved it him self. ETHICAL TANKING has been vindicated by the ardent anti tanker. T34, you are looking at the wrong dataset. The 10 win, 10 loss data set was from running bots with default loadouts: i.e. Giffins with Publisher Ts and Pins, rather than from tanking. That thread does not publish his tanking raw data, only post-processing conclusions.
|
|
|
Post by SuperHero on Jan 19, 2018 19:51:39 GMT -5
mobiuswr Your still under quarantine but your reading the wrong things into it. 1. I make no statements about "long range" being bad universally. I gave an example of two places where a zeus furry would require a very different response. With the advent of hanger decks building 2nd hanger with long and mid range is going to be almost a requirement. 2. You have drawn a spurious conclusion regarding capping hangers, and actually my own data from running them. The high win rate of a capping hanger (can be held steady in the 60%-%65 range) can make an outsized battlefield contribution - but NOT in a 5v6 where it is forced to brawl. With enough data (that we will never get) we could figure out the combat effectiveness of any given bot in any given situation, and I'm sure that we would find that things are "mostly" deterministic. Lastly, careful about what conclusions you draw from this data - it is talking about outcomes in a situation not how often those situations occur. Zeus-anything is Midrange, not long range. A beacon capping hangar hurts your side under the very likely circumstances you will have a tanker on your side in low leagues and is completely unviable in high leagues. I love how one gives data and admits it’s inconclusive or at least situational while one gives nothing and insists that his word and conclusion is gospel. keep talking. I see credibility just ebbing away.
|
|
|
Post by ᎶƦ℮℮ƊᎽ ƤΛƝƊΛ on Jan 19, 2018 20:04:24 GMT -5
Zeus-anything is Midrange, not long range. A beacon capping hangar hurts your side under the very likely circumstances you will have a tanker on your side in low leagues and is completely unviable in high leagues. I love how one gives data and admits it’s inconclusive or at least situational while one gives nothing and insists that his word and conclusion is gospel. keep talking. I see credibility just ebbing away. To be fair, neither gave data. As has been pointed out, the model is not based on data, its based on an assumption and was built with intent to produce that assumption. It's a very interesting model/analysis, just needs some work and more true data as a foundation. Been talking to zer0 on Discord about it and we can (and did) go on for hours about it.
|
|
|
Post by zer00eyz on Jan 19, 2018 21:11:07 GMT -5
I love how one gives data and admits it’s inconclusive or at least situational while one gives nothing and insists that his word and conclusion is gospel. keep talking. I see credibility just ebbing away. To be fair, neither gave data. As has been pointed out, the model is not based on data, its based on an assumption and was built with intent to produce that assumption. It's a very interesting model/analysis, just needs some work and more true data as a foundation. Been talking to zer0 on Discord about it and we can (and did) go on for hours about it. I linked to past published data (raw) that you can go and look at yourself in this thread. I will happily publish more as I gain and process it. krebby asked about models and I already highlighted what else I tried in response to hon_shu on page 1 -- Can a simple model act as an accurate predictor? It really does depend - If I'm trying to predict the weather later today, and I can see the sky, and have a barometer, and I know the season I might make a good guess. If you want to know the weather next week in some remote state then that is going to take a much more robust model. The issue is this model isn't trying to do any of those things, it's predictive powers are more or less saying "it is cold in winter, warm in summer based on where you live". As for its accuracy it is something that we can verify but ultimately is it "good enough" to make broad generalizations? Probably - ultimately all those factors that should be "in there" just keep canceling against factors that "aren't" and the end result is "random noise". As an example, I did build a model that was based on "players on the field" and then giving an advantage to the team with more of those and eliminating bots (till a player went to zero and then adjusting the advantage). One would think that this would be more representative of a game. The problem is when you start accounting for one factor (players) you need to account for for other factors their frequency of engagement, the bots, the map -- the model gets more complex... And I'm sure we could make it accurate in predicting A game assuming we knew the bots involved and the player skill and the map. It would then be terrible at making larger game predictions because we don't know about the bots of EVERY player, the SKILL of ever player, true map frequency --- and so on. At some point you can cancel this out with "random noise" -- in this case the noise is very uniform (coin flips), and it fits the data we have fairly accurately. I have already pointed out how one can easily break from this model - running a light capping hanger - something that I DONT see frequently if ever any more. I put assumptions back in to deal with that (happy to add more). Short of someone putting up more REAL DATA (results screens or summaries would be awesome) the simple model should stand. Don't just tell me I'm wrong show me how, and bring the data to say why - sadly all I have in exchange are forum upvotes.
|
|
|
Post by ᎶƦ℮℮ƊᎽ ƤΛƝƊΛ on Jan 20, 2018 1:27:00 GMT -5
zer00eyz your model has two assumptions: -All bots available are in play at all times -All game outcomes are decided by total mech out Both assumptions are incorrect, so it is mathematically impossible for the model to predict anything related to the game. You are trying to build a simple model for something that is not simple in any way. In order to build a useful model at all, you need a lot more data, and a lot more specific data, that we as players simply don't have the ability to collect. I understand the concept behind your model, and it doesn't apply in any way to the game. It's a novelty. And your data is skewed as well, because you are an active participant in every match you have data for, and you're an above average skill player. Meaning in every match you have data for, your team as an above average skill player on it. If the actual natural win rate for a 5v6 team is 20%, your higher skill rate when you're on the team of 5 will compensate for that difference. The short handed team has 80% of games that can be altered by your skill. On the flip side, if you're on the team of 6, you're far less likely to improve the odds of winning, because there are only 20% of those games that your skill can change the outcome of. That is where the 5% variance comes from. A true tanker is leaving every single game they have data for. They have no influence on the games. It's pure data from games that their team always starts down at least a man. From what I posted before, we know that average placement of other tankers will actually push win rates up slightly. So it's reasonable to say the "natural" win rate for true 5v6 matches is between 18-20%.
|
|
|
Post by mystic spastic on Jan 20, 2018 17:31:39 GMT -5
Here is some math the anti-tanker fans will love. Tankers existence pushes your win rate up. Here's the simple math, and we are assuming the matchmaker evenly(randomly) distributes tankers. In every single game you play, you are a constant. Meaning every single game you experience, you are a player in the game, and you are always on blue team. That means that in every single game you play, there are only 11 randomly assigned players. You (every single time on blue), six randomly assigned red opponents, and five randomly assigned blue teammates. Now it gets fun, throw in a tanker in any random game. There is a 5/11 chance of the tanker randomly being assigned as one of your 5 teammates. There is a 6/11 chance of the tanker being randomly assigned as one of your 6 opponents. This means that over time, the very existence of tankers pushes your win rate higher, because they almost 22% more likely to be on the other team (. 55/.45). Now I understand that tankers ruin game play regardless of which side they're on. This simply debunks the theory that tankers are hurting you personally. They're actually helping your win rate. Ironic. Save for the fact that for every tanker on the way down, there's eventually going to be a clubber on the way up. No? So, the net effect of tankers/clubbers on your overall long term win rate then becomes statistically nil. How it effects your overall sense of satisfaction with game play is what remains, overall, negative.
|
|
|
Post by ᎶƦ℮℮ƊᎽ ƤΛƝƊΛ on Jan 20, 2018 18:00:42 GMT -5
Here is some math the anti-tanker fans will love. Tankers existence pushes your win rate up. Here's the simple math, and we are assuming the matchmaker evenly(randomly) distributes tankers. In every single game you play, you are a constant. Meaning every single game you experience, you are a player in the game, and you are always on blue team. That means that in every single game you play, there are only 11 randomly assigned players. You (every single time on blue), six randomly assigned red opponents, and five randomly assigned blue teammates. Now it gets fun, throw in a tanker in any random game. There is a 5/11 chance of the tanker randomly being assigned as one of your 5 teammates. There is a 6/11 chance of the tanker being randomly assigned as one of your 6 opponents. This means that over time, the very existence of tankers pushes your win rate higher, because they almost 22% more likely to be on the other team (. 55/.45). Now I understand that tankers ruin game play regardless of which side they're on. This simply debunks the theory that tankers are hurting you personally. They're actually helping your win rate. Ironic. Save for the fact that for every tanker on the way down, there's eventually going to be a clubber on the way up. No? So, the net effect of tankers/clubbers on your overall long term win rate then becomes statistically nil. How it effects your overall sense of satisfaction with game play is what remains, overall, negative. It would depend on your league, as a clubber's affect on the outcome diminishes the higher they climb. For example, if you're in Diamond, a tanker always leaves his or her team at a disadvantage and that disadvantage can be quantified as a ~20% chance of losing. In Diamond league, even a max 12/12 clubber isn't necessarily going to win 80% of the time. So I guess essentially that the higher you are in the league system, the more "beneficial" the tanker/clubber cycle becomes to your win rate However yes, the game play diminishes no matter what. It's only the people claiming they can't win because of tankers that are way off base.
|
|
|
Post by SuperHero on Jan 20, 2018 19:48:30 GMT -5
To be fair, I don’t think most folks are upset about the win rate. I suspect that they are more upset that even when they win, they don’t get the gold payout or silver payouts or the top positions cos the tankers have taken that all. And they feel rightly or wrongly that they didn’t have a chance thanks to the tankers OP hangar
|
|
|
Post by zer00eyz on Jan 20, 2018 20:19:34 GMT -5
zer00eyz your model has two assumptions: Both assumptions are incorrect, so it is mathematically impossible for the model to predict anything related to the game. You are trying to build a simple model for something that is not simple in any way. In order to build a useful model at all, you need a lot more data, and a lot more specific data, that we as players simply don't have the ability to collect. I understand the concept behind your model, and it doesn't apply in any way to the game. It's a novelty. And your data is skewed as well, because you are an active participant in every match you have data for, and you're an above average skill player. Meaning in every match you have data for, your team as an above average skill player on it. If the actual natural win rate for a 5v6 team is 20%, your higher skill rate when you're on the team of 5 will compensate for that difference. The short handed team has 80% of games that can be altered by your skill. On the flip side, if you're on the team of 6, you're far less likely to improve the odds of winning, because there are only 20% of those games that your skill can change the outcome of. That is where the 5% variance comes from. A true tanker is leaving every single game they have data for. They have no influence on the games. It's pure data from games that their team always starts down at least a man. From what I posted before, we know that average placement of other tankers will actually push win rates up slightly. So it's reasonable to say the "natural" win rate for true 5v6 matches is between 18-20%. > -All bots available are in play at all times This is a wrong assumption - you can easily scroll up where I responded to your previous post as to why this model is likely valid. > -All game outcomes are decided by total mech out Oddly this is probably why the model WORKS for the data we have. 5v6 is a lopsided situation and more likely to result in a mechout -- we don't care WHY blue wins (or looses) just with what frequency. Were replacing a bunch of deterministic factors with random -- (how stats models work) > In order to build a useful model at all, you need a lot more data, and a lot more specific data, that we as players simply don't have the ability to collect. Right - you keep pointing to things that would be in a simulation, or a very robust model to predict very narrow outcomes. As I pointed out above, already - this model is making statement that are the equal to: "It will be cold in winter" - and your pointing the odd warm day or saying "but not in the tropics" --- > I understand the concept behind your model, See above, you completely don't understand and your statements make that clear. > and it doesn't apply in any way to the game. Actually this model lines up with the actual DATA that has been presented. > And your data is skewed as well, because you are an active participant in every match you have data for, and you're an above average skill player. Between your assumption about my skill, and the fact that the way MM works skill gets factored out (the never ending MM keeps me at %50) makes this assumption questionable at best. > A true tanker is leaving every single game they have data for. They have no influence on the games. It's pure data from games that their team always starts down at least a man. From what I posted before, we know that average placement of other tankers will actually push win rates up slightly. So it's reasonable to say the "natural" win rate for true 5v6 matches is between 18-20%. To get to the point of being able to start my current data collection project I needed to move DOWN in leagues (silver) - In order to do that I played a two bot hanger. What is odd is that my model predicts that I should have had a 33% floor in wins, and my actual floor was %30. I won't dispute the 20% - but am at a loss to explain it, we would need post match data from a tanker moving down through the league system (not just dwelling at the bottom) to see an actual floor!
|
|
Deleted
Deleted Member
Posts: 0
Karma:
|
Post by Deleted on Jan 21, 2018 0:13:05 GMT -5
... you are a player in the game, and you are always on blue team. I KNOW! I was just thinking about it, played thousands of games and always blue! What are the odds of that?! Jokes aside, great thread. Since you had to input data manually, maybe it would be worthwhile to ask Pixonic if they keep match stats?
|
|
|
Post by Dejnov on Jan 21, 2018 0:20:36 GMT -5
Under a certain set of (limited) conditions War Robots’s MM should lead to some fairly predictable outcomes if the MM is “balanced” for the most part. The reality is that outside the boundaries where resources are MOST important (the very bottom i.e. bronze and below, and the very top i.e. champs) things appear to be predictable. The question is HOW predictable? The answer might be frighteningly so. A team playing 5 members who are looking to land in the top spot have a 25% chance of wining against a full group of six. Because I lack data on “camping” (something I will be tracking in a future project) and our dear Dr’s data is accurate on the whole, lets take his information on campers at face value. What got me the results I was seeking was to simply pick a red or blue team members bot to eliminate, reducing the total number of bots available one by one. This process resulted in the magical %25 win rate. Along a similar line, I slowly began to increase the total number of bots to simulate what a camper might do (dwelling in a bot too long, and not leveraging their full hanger depth) - A camper who stays in their bot the entire match, 29% win rate, who leverages 2 bots, 35% win rate, who leverages 3 bots easily rounds up to 40% win rate. Not only do these line up nicely with the Dr’s information (I will seek to independently verify this in the coming weeks, but it will require a lot of sample data) but points to the fact that we can model games on pure chance. Notes: - I built several randomized models, and oddly coin flipping to pick a team member to eliminate fits our known data points the best
- We have a parity issue: tankers bottom out at a %20 win rate, but victory rates in 5v6’s sit at %25 - the discrepancy might be a variant of the monte hall problem or a clue to the inner workings of MM. This needs to be explored.
- This is a random model, it fits our know data well, and would indicate that MM is performing “fairly” though probably not in a sastfiying way.
- This model does not account for exceptionalism, or happiness
- We need hard numbers on games impacted by campers vs tankers to get to an overall loss rate - I will be keeping track of these things in my next sets of games to get HARD data on both and see where it fits in. I suspect that campers are "league dependent" and may have a larger overall impact than any one suspects.
I know you did the 25 v 30 to represent the tanker - did you do the 26 v 30 to represent the camper?? And did it come out to 29%? Dejnov. P.S. Between you and Dr. Yat you're changing my playstyle tremendously...
|
|
|
Post by zer00eyz on Jan 21, 2018 1:33:52 GMT -5
<snip> A camper who stays in their bot the entire match, 29% win rate, who leverages 2 bots, 35% win rate, who leverages 3 bots easily rounds up to 40% win rate. Not only do these line up nicely with the Dr’s information (I will seek to independently verify this in the coming weeks, but it will require a lot of sample data) but points to the fact that we can model games on pure chance.
I know you did the 25 v 30 to represent the tanker - did you do the 26 v 30 to represent the camper?? And did it come out to 29%? Dejnov. P.S. Between you and Dr. Yat you're changing my playstyle tremendously... It was in what you quoted me but might not be clear: 26v30: -- %29 27v30 -- 34.xxx but really 35% 28v30 --- is really 39.xxx so 40%
|
|
|
Post by frunobulax on Jan 21, 2018 4:22:51 GMT -5
To be fair, neither gave data. As has been pointed out, the model is not based on data, its based on an assumption and was built with intent to produce that assumption. It's a very interesting model/analysis, just needs some work and more true data as a foundation. Can a simple model act as an accurate predictor? It really does depend - If I'm trying to predict the weather later today, and I can see the sky, and have a barometer, and I know the season I might make a good guess. If you want to know the weather next week in some remote state then that is going to take a much more robust model. I may be a bit thick, or maybe we just have a different understanding about what a model is, but I have no clue how you come to your numbers. Could you explain that, with an example maybe? To give you an idea about what I would consider a "model": Basically, what I would do is to build a simple Markov chain model with transitions being battles for beacons. Robots would be knife fighters, cappers (cappers having less firepower) and "campers" (midrangers/snipers). I'd approximate Domination only, because it's easier. Robots can be either at a beacon or at the spawn. For initialization, create a random hangar for each player. (I'm not going into detail here. You'd need to put some thought in that, like at most 2 cappers for any given player.) To determine the influence of camping, we could run some series - say we start with one random hangar that we use for both teams, and randomly switch N cappers/knife fighters to campers for one side (N between 1 and 6 maybe). Then we observe which team has better chances to win. Players can be regular players or tankers, tankers have a hangar consisting of 1 capper robot. At any state, a Markov chain transition consists of the following steps: (1) Choose random destinations (beacons) for all robots at the spawn point. For simplicity, assume that we always have 2 near beacons and 1 middle beacons, and that robots have a high probability to choose a near beacon that is possessed by the opposing team. Also, robots will be more likely to choose a friendly beacon if the team has 3 or more beacons (defending), but more likely to choose an enemy beacon if the team has 2 or less beacons. (2) Time progress and movement: The domination bar is advanced (see below) and robot positions are changed. Robots will take damage from enemy campers if they move, depending on number of enemy campers and distance: Fewer damage if they move from spawn to a near beacon, more damage if they move from spawn to a far beacon. (3) Check for a win (domination bar is full or one team has no more robots). (4) The result of step 2 is that all robots are at one beacon and have some damage. Now we do random battle outcomes. Again no details, but we can determine a "battle strength" at each beacon depending on number and type of robots for each team and damage of the robots. (Campers affect only step 2.) The losing robots are destroyed. The winner gets the beacon, the winning robots are damaged (depending on the numbers). (5) The players that have lost robots in step (4) will respawn robots, if they still have some. (6) The result is that all beacons have an owner (let's disregard the rather small chance that a beacon is neutralized for now). For the domination bar, we have some observations about that but could get more precise assumption by observing battles. In 5vs0 beacons, a battle lasts less than 3 minutes. As beacons are not turned immediately, I'd say that it takes 2 minutes to fill the domination bar for a team leading 5-0. In 4vs1 beacon situation battles last maybe 4-5 minutes, so assume that it takes 4 minutes to fill the domination bar. In 3vs2 it takes 8 minutes to fill the domination bar. Now, if we assume that each Markov step takes half a minute we know exactly how much progress each team will get, and we may have a winner on domination bar or mechout. Obviously, there are a lot of details to be worked out, especially for steps 2 and 4 (but also for step 1). The assumptions have to be simple but reasonably close to the real thing. You mustn't be afraid to make some simplifications, too, or the whole thing will be way too complex. This way you can run some simulations. First validate your numbers - if your assumptions are correct and your code works, then you can run simulations with 6 players vs. 5 players and 1 tanker, and should arrive at a 25% win rate. (I have serious doubts about this number, but again, if that are your assumptions then go with it.) Then you can run a large number of simulations with varying initial states. This should give some interesting numbers.
|
|
|
Post by gman4hire on Jan 21, 2018 5:03:48 GMT -5
Yawn!! Just shoot some reds, have some fun and get on with your real lives people!!! Seriously the fate of the human race does not depend on the outcome of an irrelevent 10min mobile app game? The outcome of a game is not predetermined by some super computer controlling the inner thoughts and decissions of a gamer. I win less games if i play when im pissed, and im fairly sure the MM isnt buying me JD and cokes? OR IS IT??? Pass my tinfoil hat im outta here!
|
|
|
Post by frunobulax on Jan 21, 2018 5:21:41 GMT -5
Yawn!! Just shoot some reds, have some fun and get on with your real lives people!!! Seriously the fate of the human race does not depend on the outcome of an irrelevent 10min mobile app game? The outcome of a game is not predetermined by some super computer controlling the inner thoughts and decissions of a gamer. I win less games if i play when im pissed, and im fairly sure the MM isnt buying me JD and cokes? OR IS IT??? Pass my tinfoil hat im outta here! Frankly, creating wild theories in the forum is more fun than playing the game right now. "XYZ killed Frunobulax [Shocktrain Mk2]" is getting old. Nobody is forcing you at gunpoint to read it.
|
|
|
Post by gman4hire on Jan 21, 2018 6:31:21 GMT -5
I pressume most of the players collaborating in this thread are playing at the top of the tree? And I honestly feel for you i do, I appreciate that a game once domminated by skill and strategy has come down to spend a ton of cash and destroy everyone! and that it leaves a bitter taste in the mouths of the veterans that made this game what it is... or what it was. Im only 6mths in and playing D2, i dont wish to go higher and will stay in my skill lvl as long as i can. : ) That said I try and read everything you guys write as i feel wisdom is power and you guys have buckets to share... but honestly i couldnt read/understand this /\ , its all just speculative, and the shear volume of possible variants in any 1 game, means no amount of statistical analysis can determine the outcome of a game. Im just saying you all seem bright people, surely there must be a better use of your time. : )
|
|
Deleted
Deleted Member
Posts: 0
Karma:
|
Post by Deleted on Jan 21, 2018 9:35:37 GMT -5
... means no amount of statistical analysis can determine the outcome of a game. Im just saying you all seem bright people, surely there must be a better use of your time. : ) The saying among statisticians is all models are wrong but some are useful. Maybe model will not predict outcome every time correctly, but it might give you enough insight to adjust strategy. And while doing stats for War Robots might seem like a silly thing, the same skills are used in finance, engineering and etc to find most stable/optimal/ profitable solutions. I'd guess that people in this thread are data scientists in real life and this is what these people do for fun. Regards
|
|
|
Post by zer00eyz on Jan 21, 2018 12:56:27 GMT -5
I may be a bit thick, or maybe we just have a different understanding about what a model is, but I have no clue how you come to your numbers. Could you explain that, with an example maybe? To give you an idea about what I would consider a "model": Basically, what I would do is to build a simple Markov chain model with transitions being battles for beacons. Robots would be knife fighters, cappers (cappers having less firepower) and "campers" (midrangers/snipers). I'd approximate Domination only, because it's easier. Robots can be either at a beacon or at the spawn. For initialization, create a random hangar for each player. (I'm not going into detail here. You'd need to put some thought in that, like at most 2 cappers for any given player.) To determine the influence of camping, we could run some series - say we start with one random hangar that we use for both teams, and randomly switch N cappers/knife fighters to campers for one side (N between 1 and 6 maybe). Then we observe which team has better chances to win. Players can be regular players or tankers, tankers have a hangar consisting of 1 capper robot. At any state, a Markov chain transition consists of the following steps: (1) Choose random destinations (beacons) for all robots at the spawn point. For simplicity, assume that we always have 2 near beacons and 1 middle beacons, and that robots have a high probability to choose a near beacon that is possessed by the opposing team. Also, robots will be more likely to choose a friendly beacon if the team has 3 or more beacons (defending), but more likely to choose an enemy beacon if the team has 2 or less beacons. (2) Time progress and movement: The domination bar is advanced (see below) and robot positions are changed. Robots will take damage from enemy campers if they move, depending on number of enemy campers and distance: Fewer damage if they move from spawn to a near beacon, more damage if they move from spawn to a far beacon. (3) Check for a win (domination bar is full or one team has no more robots). (4) The result of step 2 is that all robots are at one beacon and have some damage. Now we do random battle outcomes. Again no details, but we can determine a "battle strength" at each beacon depending on number and type of robots for each team and damage of the robots. (Campers affect only step 2.) The losing robots are destroyed. The winner gets the beacon, the winning robots are damaged (depending on the numbers). (5) The players that have lost robots in step (4) will respawn robots, if they still have some. (6) The result is that all beacons have an owner (let's disregard the rather small chance that a beacon is neutralized for now). For the domination bar, we have some observations about that but could get more precise assumption by observing battles. In 5vs0 beacons, a battle lasts less than 3 minutes. As beacons are not turned immediately, I'd say that it takes 2 minutes to fill the domination bar for a team leading 5-0. In 4vs1 beacon situation battles last maybe 4-5 minutes, so assume that it takes 4 minutes to fill the domination bar. In 3vs2 it takes 8 minutes to fill the domination bar. Now, if we assume that each Markov step takes half a minute we know exactly how much progress each team will get, and we may have a winner on domination bar or mechout. Obviously, there are a lot of details to be worked out, especially for steps 2 and 4 (but also for step 1). The assumptions have to be simple but reasonably close to the real thing. You mustn't be afraid to make some simplifications, too, or the whole thing will be way too complex. This way you can run some simulations. First validate your numbers - if your assumptions are correct and your code works, then you can run simulations with 6 players vs. 5 players and 1 tanker, and should arrive at a 25% win rate. (I have serious doubts about this number, but again, if that are your assumptions then go with it.) Then you can run a large number of simulations with varying initial states. This should give some interesting numbers. You forgot, power (wepon and bot), speed, skill (this is a whole sub-chain), teamwork (low or high), device used, connectivity (drops, lag) --- To your point the list here is huge. It would also be hard to get the right parameters around all of these but at the end I would be able to tell you the "odds" of you winning a fight in a certain set of circumstances... and what you might be able to do to change those conditions... That wasn't my goal - it was to build the simplest model possible. The fact that the data we have maps close enough to a non deterministic system that boils down to a coin flip is telling. Sure there are factors that ONE player can control, but are they enough to swing the fight is something that we might be able to determine with a Markov chain, but the reality is that this is a good enough approximation. It should drive FURTHER data collection, that might lead to a more robust (but still very simple) model. > ...should arrive at a 25% win rate. (I have serious doubts about this number, but again, if that are your assumptions then go with it.) Im curious WHY you think this is off base? And what parameters your placing around that judgement... intuition is sometimes telling.
|
|
|
Post by ᎶƦ℮℮ƊᎽ ƤΛƝƊΛ on Jan 21, 2018 13:07:15 GMT -5
You picked an outcome and sat there adjusting your model until it spit your outcome, and the parameters you settled on (again, only because they spit out the desired results) are not applicable to War Robots. You can build a model with intent to come to a predetermined outcome and then say it's a good model.
I could come up with a laundry list of models that will spit out 25% and use some combination of a 5 to 6 ratio. That doesn't mean it's a model, it's a novelty simulator.
|
|
|
Post by frunobulax on Jan 21, 2018 15:23:48 GMT -5
You forgot, power (wepon and bot), speed, skill (this is a whole sub-chain), teamwork (low or high), device used, connectivity (drops, lag) --- To your point the list here is huge. It would also be hard to get the right parameters around all of these but at the end I would be able to tell you the "odds" of you winning a fight in a certain set of circumstances... and what you might be able to do to change those conditions... Well, that's the issue with simplification I would abstract from all that, because if you go there, you'd also need knowledge about the matchmaking and how the matchmaking pits perceived strong and weak players. See war-robots-forum.freeforums.net/post/243954/thread. > ...should arrive at a 25% win rate. (I have serious doubts about this number, but again, if that are your assumptions then go with it.) Im curious WHY you think this is off base? And what parameters your placing around that judgement... intuition is sometimes telling. There are two things to consider. (a) Does the matchmaking try to compensate for tankers? The general assumption seems to be that the matchmaking is impartial and uses only the league position and possibly the win rate, while I have my doubts about that (see link above). Let's say you observe 100 battles with 1 tanker, and the non-tanking side wins 25 (%) of the battles. Does that mean that the presence of the tanker lowers the chances from 50% to 25%? This is not necessarily true. For example, if the matchmaking would try to compensate for the tanking, that is, it would detect the tanker as "weak" and tries to match him with "strong" players, then we might have 5 "strong" players and a tanker playing against 6 normal players. If that is the case, we would have a lower winning rate if the tanker would be placed in a team with random (and not strong) players. That matches my intuition to some degree. Superiority in numbers is so crucial in War Robots: Unless we have some outlandish builds or really bad players, a 2vs1 situation will usually keep both robots standing, even though one might be damaged. Even if the 5-robot team wins one matchup, it should lose either 2 other matchups or lose beacons in the process. And it is just a gut feeling, but I'd say that numerical superiority is almost impossible to overcome, unless we have a fader/tanker on the 6-robot team. Soccer is big in my home country, but if 1 player is disqualified for one team (unlike Football, where an ejected player is simply replaced, the team must play with a player less), a 10vs11 situation leads to a loss in over 80% of the cases, and a win is nearly unheard of. And there are no ties in War Robots (b) Do players react to it? Like, if I see a player ejecting 4 bots and jumping into the fray with an Ecu Cossack, am I more likely to quit playing myself? This could affect numbers significantly too, but in the other direction.
|
|
|
Post by Dr. Yat on Jan 31, 2018 10:06:02 GMT -5
Wow - some VERY serious math going on! My claim about the 25% win rate is based on observation, not a good mathematical model. I've been able to see this for myself and on tanker profiles. For me, it was the very first event (maybe the fourth anniversary) that had a task chain. I was (sorry, all) grabbing one or two beacons and then ditching the match. After a long chain of this, having left my team in a five vs. six situation, I had a 24% win rate. And that number appears a LOT when you see individual tankers doing their thing. When they team tank, it approaches zero, because the team with no members ALWAYS loses!!
|
|