Suggestion to improve the alliance choosing program

[Bongle]

Originally Posted by XaulZan11

I understand what you are saying, but if you talk to nearly every psycologist, past behavior is the best predictor of future behavior. If there was no correlation or indication, then we would expect teams like 71, 111, 233, 1114 to have a normal distribution of results (ie win 3 regionals in a year just as often as not getting picked for the eliminations in one year). As we know, however, these teams always are some of the top teams. I think you mean that the correlation is not strong enough to be a used. If so, I agree.

Again to go back to the stats I did a few years back, a team's seeding performance in one year is a very poor predictor of its performance in the year following.

Here's a graph:http://www.chiefdelphi.com/forums/at...6&d=1175831673

On the X axis is a team's seeding performance in 2005. Further left is better. On the Y axis is a team's seeding performance in 2006, lower is better. You'll see about the only thing you can predict is that teams who were top seeds in 2005 tended to not be dead last in 2006. Likewise, teams who did very poorly in 2006 tended to not win the following year (but some did). Past behavior predicting future behavior may work well in humans, but not so much in robotics teams.

The teams that do well year after year are very special cases. Out of the 1500ish active teams in FIRST, people can probably only name 50ish 'power houses' who win year after year after year and never hiccup.


On the main topic:
Keep in mind that a team's next-year performance will probably be modified FAR more by who they communicate in the pits with, rather than who they play with on the field. If you play 8 games, you're only on the field for 16-20 minutes, but you're at the regional in the proximity of other robotics teams (whether in the hotel, pits, stands, or fields) for 72 hours. The 71 hours and 40 minutes that you're not on the field are where your entire team can learn from vets, not just on-field.

[Kimberly]

There seems to be some confusion about what I actually mean when I say it would be better to have the rookie teams more evenly distributed on alliances.

What I am NOT saying is rookies are worse than experienced teams.

What I AM saying is rookies are less experienced than experienced teams.

Sometimes the student is better than the teacher, but that doesn't mean the student can't still learn from the teacher because experience does count for something.

Having said that though, if the data provided is correct, there does seem to be a correlation between the number of rookie teams on an alliance and how the alliance performs, so if it's possible to reduce the "unfairness" to the non-rookie teams by even a little bit through distributing the rookie teams more evenly, why in the world would anyone not want to?

Excepting, of course, the difficulty in writing the program that can do this and all the other functions too. A valid and reasonable argument against my suggestions, btw. If good programmers say it's too hard to factor rookies into the algorithm and still have it work as well as it does now, then it's too hard.

[GaryVoshol]

Kimberly may have a point - they got a really unlucky schedule at GLR. In their first 4 matches, they were paired with another rookie. They were paired with a "powerhouse" veteran team just once, 469 - and in that match they came up against 494 and 67.

To contrast, I looked at Rush's schedule (lowest team number at GLR). They saw rookies on the field only twice - in the same match, one on each alliance. 33 played with a rookie once, never against one.

503, in the middle of the list of team numbers, played with and against rookies and very low-number veterans in many of their matches. 573 had a similar schedule.

2676, the highest team number at GLR, played with and against rookies in most of their schedule.

I looked at the West Michgan schedule and saw a similar clustering of rookie teams, and again of low-number teams.

We've seen that the "maximize time between matches" constraint often results in a team playing with another team in one match and against them the following. That's not unreasonable.

Does the schedule algorithm shuffle the teams before slotting them into the schedule? Or does it start in numeric order? As we saw last year, strict team order doesn't equate to team strength. But did the schedule inadvertently create semi-tiers?

[Bongle]

Originally Posted by GaryVoshol

...

Hmmm... I've even seen griping from a low-numbered team (I think it was 48) mentioning that they were in the first match of a regional multiple times.

Wait 10 minutes, I'll re-jigger my OPR calculator to get an idea of the average team # that a team plays with.

[XaulZan11]

Originally Posted by Bongle

Again to go back to the stats I did a few years back, a team's seeding performance in one year is a very poor predictor of its performance in the year following.

Here's a graph:http://www.chiefdelphi.com/forums/at...6&d=1175831673

On the X axis is a team's seeding performance in 2005. Further left is better. On the Y axis is a team's seeding performance in 2006, lower is better. You'll see about the only thing you can predict is that teams who were top seeds in 2005 tended to not be dead last in 2006. Likewise, teams who did very poorly in 2006 tended to not win the following year (but some did). Past behavior predicting future behavior may work well in humans, but not so much in robotics teams.

The teams that do well year after year are very special cases. Out of the 1500ish active teams in FIRST, people can probably only name 50ish 'power houses' who win year after year after year and never hiccup.

I never said it was a fantastic predictor. I said it was the best. Predicting behavior in humans is very challenging.

The graph that you provide may very well prove that past preformance does predict future behavior. It appears the correlation coefficent would be around .2 or .3 and with a large sample size (around 1,000) I would not be surprised to see the correlation statistically significant. This means that the observed relationship is not due to random varation but because the two samples (results from 2005 and 2006) are in fact related.

Although it is not a perfect correlation or relationship (nearly all relationships in life are not perfect) doesn't mean there is none. It may not be a very strong relationship, but it appears there is one. I cannot think of another relationship (team number, funding...) that is a better predictor than past preformance.


To tie this back into the original topic, there is not perfect predictor for team preformance. Unless someone does a huge multiple regression study and finds a way to predict how teams are (I don't think there is one), the best way is to just randomly assign teams under the perameters (such as time inbetween matches) to ensure the most fair pairings.

[Bongle]

It does appear that lower-numbered teams had lower-numbered partners.

Using all the week 2-5 regional results, I found the average team number of each team's alliance partners. The results:

0-numbered teams played with avg team nbrs of 1091, against avg nbrs 1080
100-numbered teams played with avg team nbrs of 1093, against avg nbrs 1095
200-numbered teams played with avg team nbrs of 1073, against avg nbrs 1076
300-numbered teams played with avg team nbrs of 1083, against avg nbrs 1061
400-numbered teams played with avg team nbrs of 1158, against avg nbrs 1157
500-numbered teams played with avg team nbrs of 1147, against avg nbrs 1200
600-numbered teams played with avg team nbrs of 1264, against avg nbrs 1253
700-numbered teams played with avg team nbrs of 1269, against avg nbrs 1192
800-numbered teams played with avg team nbrs of 1268, against avg nbrs 1249
900-numbered teams played with avg team nbrs of 1351, against avg nbrs 1392
1000-numbered teams played with avg team nbrs of 1290, against avg nbrs 1314
1100-numbered teams played with avg team nbrs of 1382, against avg nbrs 1372
1200-numbered teams played with avg team nbrs of 1398, against avg nbrs 1357
1300-numbered teams played with avg team nbrs of 1399, against avg nbrs 1436
1400-numbered teams played with avg team nbrs of 1376, against avg nbrs 1403
1500-numbered teams played with avg team nbrs of 1394, against avg nbrs 1393
1600-numbered teams played with avg team nbrs of 1461, against avg nbrs 1452
1700-numbered teams played with avg team nbrs of 1446, against avg nbrs 1444
1800-numbered teams played with avg team nbrs of 1556, against avg nbrs 1533
1900-numbered teams played with avg team nbrs of 1687, against avg nbrs 1678
2000-numbered teams played with avg team nbrs of 1522, against avg nbrs 1528
2100-numbered teams played with avg team nbrs of 1563, against avg nbrs 1574
2200-numbered teams played with avg team nbrs of 1760, against avg nbrs 1730
2300-numbered teams played with avg team nbrs of 1757, against avg nbrs 1762
2400-numbered teams played with avg team nbrs of 1713, against avg nbrs 1735
2500-numbered teams played with avg team nbrs of 1793, against avg nbrs 1799
2600-numbered teams played with avg team nbrs of 1630, against avg nbrs 1630

I'm fairly sure that unless I have a bug, this indicates that higher-numbered teams indeed don't get to play with lower-numbered teams as often as randomness would suggest, which implies that the match-making algorithm is not blind to team numbers like I would think it should be.


Here's the summing code. I grouped things by hundreds so patterns would show up and it'd be easier to implement. m is a match object, which includes the team numbers of the red and blue alliances. m.red and m.blue are 3-element integer arrays of the alliance. If I have a big glaring bug, please point it out.

for(int x = 0;x < lstMatches.size();x++)
	{
		Match m = lstMatches[x];

		iTeamWithSum[m.blue[0] / 100] += m.blue[1] + m.blue[2];
		iTeamWithSum[m.blue[1] / 100] += m.blue[0] + m.blue[2];
		iTeamWithSum[m.blue[2] / 100] += m.blue[0] + m.blue[1];
		iTeamWithCount[m.blue[0] / 100]+=2;
		iTeamWithCount[m.blue[1] / 100]+=2;
		iTeamWithCount[m.blue[2] / 100]+=2;

		iTeamWithSum[m.red[0] / 100] += m.red[1] + m.red[2];
		iTeamWithSum[m.red[1] / 100] += m.red[0] + m.red[2];
		iTeamWithSum[m.red[2] / 100] += m.red[0] + m.red[1];
		iTeamWithCount[m.red[0] / 100]+=2;
		iTeamWithCount[m.red[1] / 100]+=2;
		iTeamWithCount[m.red[2] / 100]+=2;

	}
	for(int x = 0;x < iTeamsMax;x++)
	{
		cout<<(x*100)<<"-numbered teams played with avg team nbrs of "<<iTeamWithSum[x] / iTeamWithCount[x]<<endl;
	}

[XaulZan11]

Assuming those numbers Bongle provided are correct, that is a problem. I haven't completely gone through how the pairings are assigned, but it appears the team number (or the order of the teams) is a determining factor in the pairings (if not, we wouldn't have the relationship that Bongle provided). Would it be possible to randomly assign each team a number? Like say there are 50 teams at a regional, randomly assign each team a number from 1-50. Then, after the program assigns the pairings, replace the number (1-50) with the team number.

[GaryVoshol]

Using raw team numbers as integers does not create a fair test. Some regionals have very low team numbers, with only a few rookies (Detroit). Some have very high team numbers, mostly rookies (Hawaii, Minnesota, Oklahoma). The two can't be equated directly; they skew the distribution. You would have to figure out a way to assign an equivalent rank to each team number in each regional.

[SteveJanesch]

Originally Posted by Bongle

Here's the summing code. I grouped things by hundreds so patterns would show up and it'd be easier to implement. m is a match object, which includes the team numbers of the red and blue alliances. m.red and m.blue are 3-element integer arrays of the alliance. If I have a big glaring bug, please point it out.

Bongle,

I'm not sure, but I think your code includes the numbers of all three teams on each alliance. If you're calculating the average team number of only of the teams you're paired with, wouldn't you want to keep your own team number out of the average? If that's the case, it'll flatten out the "who you are paired with" average with respect to your own team number. It won't change your opponent's average team number, though. I could easily be reading the code wrong. I thought what was important was the average team number of who you were paired with, which should ideally be constant versus team number, not the whole alliance, which your own team number will influence.

- Steve

[Bongle]

Originally Posted by stj_1533

Bongle,

I'm not sure, but I think your code includes the numbers of all three teams on each alliance. If you're calculating the average team number of only of the teams you're paired with, wouldn't you want to keep your own team number out of the average? If that's the case, it'll flatten out the "who you are paired with" average with respect to your own team number. It won't change your opponent's average team number, though. I could easily be reading the code wrong. I thought what was important was the average team number of who you were paired with, which should ideally be constant versus team number, not the whole alliance, which your own team number will influence.

- Steve

Well, when it's doing the summing for the team's you're with, it leaves your team out of your bucket. iTeamWithSum is an array of 26 elements. iTeamWithSum[0] represents the sum of all alliance partners of (0-99) numbered teams, iTeamWithSum[1] is the sum of all alliance partners of (100-199) numbered teams, and so on. So to say that "188 (a 100-199 team) had team numbers 865 and 703", the code would be iTeamWithSum[1] += 865 + 703.

Example: Let's say we have an alliance of 1114, 1503, and 1680.

It would do:
iTeamWithSum[11] += 1680 + 1503 // 1114 was paired with 1680 and 1503
iTeamWithSum[15] += 1114 + 1680 // 1503 was paired with 1114 and 1680
iTeamWithSum[16] += 1114 + 1503 // 1680 was paired with 1114 and 1503

average[11] = (1680 + 1503) / 2; // note that this isn't affect by 1114 being on the 1114/1503/1680 alliance

So the actual team number that we're doing the indexing by is not included in the average, and thus shouldn't effect the final average (for that bucket).

Good argument though, keep them coming. We don't want to make a fuss if my code is simply wrong.

[SteveJanesch]

Originally Posted by Bongle

So the actual team number that we're doing the indexing by is not included in the average, and thus shouldn't effect the final average (for that bucket).

That's what I was suggesting, I'm just not sharp enough this afternoon to pick it out of the code. Well done.

[Kims Robot]

Originally Posted by Kimberly

What I am NOT saying is rookies are worse than experienced teams.

What I AM saying is rookies are less experienced than experienced teams.

I see what you are saying in general, but Im not sure this would work for long, or even in a lot of cases. As FIRST grows and has now been around since 1992, there are more and more teams being started with mentoring teams, more rookie teams attending preseason events, more rookie teams being founded by FIRST college graduates or past FIRSTers that have moved, so the level of the rookie teams has drastically increased from years ago. There are also more and more rookies going to multiple events, at which point they may have enough experience to perform more like a veteran team. The other factor is there are many older teams that go through upheaval, and either loose sponsorship for a year, come back into being, or just have a complete switch of students and/or mentors. At any of these points, the lower number team may "act" like a rookie, and in some of the ones I mentioned earlier, the rookie team may "act" like a veteran. The year that our team was a rookie team, I was amazed with how much the rookies in our area had done and how much they appeared like veteran teams, especially on the field.

While I understand what you mean, especially in certain regions, I dont think this would work as well overall, based on what we have seen with past match algorithms based on number, and the level of many of the rookie teams these days.

[StevenB]

Originally Posted by Bongle

It does appear that lower-numbered teams had lower-numbered partners. I'm fairly sure that unless I have a bug, this indicates that higher-numbered teams indeed don't get to play with lower-numbered teams as often as randomness would suggest, which implies that the match-making algorithm is not blind to team numbers like I would think it should be.

I don't think this is the case. As Gary pointed out,

Originally Posted by GaryVoshol

Some regionals have very low team numbers, with only a few rookies (Detroit). Some have very high team numbers, mostly rookies (Hawaii, Minnesota, Oklahoma).

It might be best to do this on a regional-by-regional basis, and see what you get. For example, the average team number at OKC was 2077. Almost everyone had a high team number, and thus their partners also had high numbers.

[Rick TYler]

Originally Posted by StevenB

It might be best to do this on a regional-by-regional basis, and see what you get. For example, the average team number at OKC was 2077. Almost everyone had a high team number, and thus their partners also had high numbers.

Perhaps you could normalize team numbers by representing team numbers as plus or minus from the median (mean?) team number at the regional. Or even use ordinals of a team-number stacking at each regional. Isn't statistics fun?

[Bongle]

Ok, I'll do another set of these once I have time. Here is my proposal:

On a per-regional basis:
For each team, take the average team # of all their allied teams. Compare this with the average team # at the regional (minus themselves of course) to represent their expected average team # of their allied teams if they were paired randomly. Call the difference D, and say that D for a given team is given by D = f[t]. Sp if you had a regional with teams (1,2,3,4,5,6,7,8,9,10) and 1 was paired with 2 and 3 for a match and 4 and 5 for another, D = ((2+3+4+5+6+7+8+9+10)/9 - (2+3+4+5)/4).

Hypotheses:If the scheduler is blind to team numbers, then f should not follow any pattern as t increases. If the scheduler is affected by team numbers and there is clustering, then f should follow some pattern as t goes from low team numbers to high team numbers.