If genetic fitness actually matters, is life limit still necessary?

jasonrohrer · 2019-10-30 16:18:26

The life limit was put in place long ago to make each life precious and curb rampant suicide.

However, in its current form, it mostly exerts psychological pressure, because people don't like to watch a number going down. But 12 lives with 3+ lives an hour is plenty for most people, and they will never run out. And even frequent suicide won't eat into the pool that much.

I was never 100% satisfied with that as a solution to make people care about life, but it was a stop-gap, and it mostly worked. We no longer have thousands of baby suicides per day, etc.

Genetic fitness was introduced later to make you care about your offspring. In its initial form, it mostly exerted psychological pressure, with a number going up and down. The bonus for a high score (big stomach in old age) wasn't that enticing to people. So a good portion of the player population ignored it. That was okay, because the goal was to have SOME players behave like really caring mothers, and it worked.

Everybody (I'm guessing) cares about tool slots. So everybody suddenly cares about genetic scores. Even the would-be serial-suicide-baby will be put off by FINALLY getting born into the desired situation, but having far fewer tool slots than they used to have.

In this update, I've added messaging to every life, explaining your current score and how many bonus slots you get. So it will be clear to everyone what happens when that number goes up and down.

With all that in place, is it time to retire the life limit?

Maybe there's no harm to keeping it in there as a rear firewall.

I noticed that during the recent baby apocalypse, people were equally miffed about running out of lives and taking a huge hit to their genetic score.

Anyway, I'm hoping that people will eventually care about their offspring enough that it will be the primary motivator for everything they do in the game. It should by WHY they want to keep civilization going, even when oil runs out, etc. It should by why they want to migrate (in a post-rift world), instead of just opting to sigh and die. Because in real life, this is pretty much it in terms of WHY.

I don't know that tool slots are enough.... but it feels like a much stronger motivator than what we had before.

I'm reminded of the movie The Impossible:

https://www.youtube.com/watch?v=Bgw394ZKsis

Thaulos · 2019-10-30 16:31:40

Yes it is. Most people don't care bout the number at all.

The number as stands right now it's pretty useless itself. One bad life and you go from 50 to 30 or even lower. And even if it was perfectly balanced and with no exploits most people would still not care about it at all.

Meanwhile you might want to reduce the jump in score on every life. Maybe even dividing the score increase/decrease by 10 could be better to avoid variance. With scores varying to wildly no one is going to invest into keeping high score. Why bother when you can put back from bottom to top 10 in a couple lives.

Last edited by Thaulos (2019-10-30 16:31:57)

Dodge · 2019-10-30 16:34:04

Certain groups of players do not care at all about having more tools slots (griefers, roleplayers, others) and will be content with only being able to use 2-3 tools (knife, bow +other), these players will keep suiciding either to choose the life they want or simply to annoy other players.

I can already imagine the /die spam and baby bones everywhere.

jcwilk · 2019-10-30 16:48:16

Dodge wrote:

Certain groups of players do not care at all about having more tools slots (griefers, roleplayers, others) and will be content with only being able to use 2-3 tools (knife, bow +other), these players will keep suiciding either to choose the life they want or simply to annoy other players.
I can already imagine the /die spam and baby bones everywhere.

As someone who mostly plays this game out of curiosity in the direction it's going and for exploring and nomading, and occasionally stirring up some shit and doesn't really want to do work in a game, can confirm. 2 tool slots is plenty for a little minimum effort help around the village and basic nomading.

I think dangling being an eve for really high performers would be a really strong motivator. Even for people who just like causing trouble the draw of being an eve is strong since it means you have the chance to start a new troublemaking family, and it would probably mean that they'd actually behave more so that they could cause a higher tier of chaos later ?

Amon · 2019-10-30 16:59:57

The number stopped being an issue the moment you were able to bank in a lot more lives than one per hour.
I'd say yes, let it go.
It's long overdue. there's better pressures at stake nowdays.

jasonrohrer · 2019-10-30 17:19:54

Thaulos, any thoughts on reducing variance with smaller jumps while also not increasing grind on the way up?

The jump from 50 down to 30 in one life can only come from having almost all your offspring die young. How often are people experiencing this? Do we have any screen shots of someone's genetic history screen showing such a huge drop?

Likewise, you could jump UP from 30 to 50 in one life, right? By having lots of children that all live to old age? If you're at 30, a child who lives to 60 is worth 3 points.

But yeah, if you're at 50, a baby that dies young docks you 5 points.

Part of the problem is that once you get a high score, you can pretty much only go down. That's a side-effect of keeping scores bounded. In the above example, you can see that you drop by 5 points per baby on your way down from 50, but you only rise by 3 points per old-age-death on your way back up.

Also, if you're up at 50 and you have every other baby live and die, you will drop back down to 30 over time. In that way, it's kind of a moving average.

If 1/10 of your BB dies at 0 (and the rest live to 60), your actual average is 54.

If you're at 54, and you have 9 BB live to 60, that takes you up to 57.675477066

Then if you have 1 BB die at 0, that will take you down to 51.9079293594

I ran a little sim of this happening 10 times (over 100 BB):

So it does look like it goes down and down over time. However, it eventually converges. Running this 10,000 times (for 100,000 BB) we get this as the last cycle:

And like magic, if we take the average of these last 10 values, we get 54.0000

So this is working, to effectively make your genetic score converge to the actual average of your recent offspring's lifespans.

Thaulos · 2019-10-30 17:35:12

I would guess a quick way of making sure variance isn't too bad but while still providing a quick "startup" would be to somehow factor in the number of lives you previously played.

So maybe first life it would be a big jump, second a big but smaller one, etc. Until you cap variance at the average of your past 10 lives (as an example).

For the score to mean anything it should be hard to drop and to climb. One should only really have high scores if they are consistent across their lives.

On another note, should we really evaluate genetic score based on whether or not your children lives up to 60? When does it stop being an evaluation of your skill as a mother/brother and starts becoming random luck? Maybe you should only be responsible for your children lives up until a certain age, then it's up to them to survive. What if they don't want to waste food and die at 58. Is that worse than eating that food and surviving two more years to 60?

Maybe genetic fitness could last up until your children are 20 years old?

Another thought is that wouldn't having 2 children dying at 40 be better than one children surviving to 60? Currently the score would penalize the 2 children but genes are already propagated by the time they are forty. Some players suicide at 40 to start again another live. Should their mothers be penalized for that?

Last edited by Thaulos (2019-10-30 17:38:31)

jasonrohrer · 2019-10-30 17:47:09

Yes, that's a good point about 40 vs 60. Maybe it should be adjusted to stop counting after 40, and living after 40 should be gravy?

Here's a graph of the fitness scores of the four players who got the four highest recent scores:

Note that these lines do NOT happen concurrently in time. They each represent totally different time spans.

It's hard to stay at the top for very long, and precipitous drops are indeed common for these top players.

That purple line wants to stay at the top SOOOO bad.... but it eventually falls.

Let's look at a recent portion of the purple line:

Tomorrow_Dyke	45.1737 years old	Little_Brother	31.4837	+ 1.369	32.8527	2019-10-02 08:07:06
Han_Dyke	15.8745 years old	Nephew	33.218	- 1.7343	31.4837	2019-10-02 08:04:34
Joon_Dyke	17.7918 years old	Nephew	34.932	- 1.714	33.218	2019-10-02 08:04:10
Hosea_Dyke	47.904 years old	Little_Sister	33.4907	+ 1.4413	34.932	2019-10-02 08:03:07
Stryker_Dyke	23.6717 years old	Nephew	34.5817	- 1.091	33.4907	2019-10-02 07:59:02
Nnamdi_Dyke	31.0547 years old	Little_Sister	34.9736	- 0.3919	34.5817	2019-10-02 07:56:59
Easton_Dyke	37.4641 years old	Little_Brother	34.6969	+ 0.2767	34.9736	2019-10-02 07:54:52
Savior_Dyke	36.1902 years old	You	34.531	+ 0.1659	34.6969	2019-10-02 07:51:00
Nameless	11.0639 years old	Little_Brother	37.1385	- 2.6075	34.531	2019-10-02 07:46:09
Ewen_Dyke	2.4694 years old	Little_Brother	40.9906	- 3.8521	37.1385	2019-10-02 07:23:38
Carrie_Dyke	13.1915 years old	You	44.0794	- 3.0888	40.9906	2019-10-02 07:14:43
Nameless	0.5411 years old	You	48.917	- 4.8376	44.0794	2019-10-02 07:00:56
Blessing_Raine	60 years old	You	47.6856	+ 1.2314	48.917	2019-09-24 06:10:12
Berry_Raine	27.8667 years old	Daughter	49.8877	- 2.2021	47.6856	2019-09-24 06:06:16
Ember_Fire	13.6213 years old	You	53.9173	- 4.0296	49.8877	2019-09-24 05:10:03
Sapphire_Audette	60 years old	Grandson	53.2414	+ 0.6759	53.9173	2019-09-23 14:00:08
Diamond_Audette	57.7872 years old	Granddaughter	52.7363	+ 0.5051	53.2414	2019-09-23 13:52:21
George_Audette	60 years old	Son	51.9292	+ 0.8071	52.7363	2019-09-23 13:34:04
Mary_Audette	60 years old	Daughter	51.0325	+ 0.8967	51.9292	2019-09-23 13:30:28
Winner_Audette	60 years old	Daughter	50.0361	+ 0.9964	51.0325	2019-09-23 13:20:53

This is a little messy in terms of formatting. Sorry about that. But look at Carrie_Dyke. That's the purple line player themselves, and they starved to death at age 13. So.... what are we supposed to do with that? It must cause a substantial drop in score, right?

The other option is to factor in some kind of average, or "guessed" average. Let's say we pick 30. Then lives lived over 30 are positive points, and lives that die under 30 are negative points. But can your score grow infinitely large over time then?

jcwilk · 2019-10-30 17:47:41

jasonrohrer wrote:

So it does look like it goes down and down over time. However, it eventually converges. Running this 10,000 times (for 100,000 BB) we get this as the last cycle:
51.709164
52.538248
53.284423
53.955981
54.560383
55.104344
55.593910
56.034519
56.431067
50.787960
And like magic, if we take the average of these last 10 values, we get 54.0000
So this is working, to effectively make your genetic score converge to the actual average of your recent offspring's lifespans.

6 points seems like a huge change for one baby dying from popping out while in the middle of getting griefed or something like that. Maybe it would be worth being really clear about what the goal of the algorithm is?

Do we want it to be possible for someone to remain #1 through being consistently better than everyone else at mothering? Or do we want randomness to give #2-5 brief moments in the sun at the #1 spot? Or does the #1 spot not matter as much as being at least in the top 10? Or does it only matter if they have a score of at least 50? Once the goals are clearly defined then it's not difficult to work backwards to what the math should be.

But yeah looking at things like how chess ratings work and the math behind how they do diminishing return over time might provide some inspiration and give a little more meaning to success rather than knowing it's short lived until you get born to a starving nomad and 4 kids pop out and you heroically manage to keep 3 alive but the 4th one destroys a week of progress and puts you back below a dozen other players who kill babies daily (contrived example, maybe daily/hourly rather than weekly/daily fits the described simulations better)

Coconut Fruit · 2019-10-30 17:47:49

Is life limit still necessary? Don't remove it yet, you just released tool system, give it some time.
Not everyone cares yet.

Another, probably rare case would be that someone could make a simple bot that would keep /dying, just to make lots of trash (baby bones). Probably it would be cursed out eventually, but not necessarily.

You probably don't want people run out of lives when next baby apocalypse happens, maybe it would be the best to somehow automatically detect such thing and then when it happens temporarily disabling life limit.

jasonrohrer · 2019-10-30 18:17:15

jcwilk, the existing algorithm is indeed modeled very closely on Chess Elo rankings.

How much your score jumps after a game of Chess depends on how surprising the result is. If a grandmaster is beaten by the novice, that's a very surprising result, and the novice's score jumps a lot, and the grandmaster's score drops a lot. However, if one grandmaster beats another, that's not a huge surprise, so their scores barely change. The winner goes up a tiny bit, an the loser goes down a tiny bit. Elo has the nice property of converging over time in such a way that it can be used to estimate chances of winning against a given opponent. If we have equal Elo, we are expected to beat each other 50% of the time. A grandmaster (2800) is expected to beat a novice (700) 99.9994% of the time.

How much your score jumps after an offspring dies in OHOL depends on how surprising the result is. Given that your OHOL genetic score is a moving estimate of the average lifespan of your offspring so far, if an offspring exceeds that averge, your score goes up. If an offspring dies much younger than that average, your score goes down.

The top players in OHOL with high scores are like grandmasters in chess. They have nowhere to go but down. If your Elo is 2800, there's no one you can play that will make your ranking go up very much, and a huge number of people you can play (everyone) who will make your ranking go down if you lose to them.

So if your genetic score is 55, that means that on average, you and your offspring have been living 55 years. Living to 60 isn't that surprising, so your score only goes up a little when an offspring lives to 60. But dying at 0 is a HUGE surprise, so your score drops a lot when that happens.

What if they don't want to waste food and die at 58

If you have a high enough genetic score such that dying at 58 will hurt you, then you live to old age with a large stomach, and don't waste any food.... However, it's true that the genetic fitness of your offspring has nothing to do with you, so who knows what their old-age stomachs are....

We could start matching BB with the mother who has the closest genetic score... like attracts like...

Though novice players should probably be raised by experts, so that's probably a bad idea.

Coconut Fruit · 2019-10-30 18:35:44

As for genetic fitness I would suggest doing these things:
- Leave it like it is, but make that we get maximum number of additional slots when our genes are above 40.0
- lower the amount of additional skill slots that we can get from genetics to max 2
- add another stat: your average life span past 10 lives
- make it that we get additional skill slot for living 50+ years on average or 2 slots for living 55+ years on average.

This would give us max 4 additional skill slots.

There could be also 5th additional skill slot for actually working in a town, maybe every use of our learned tools would have some value... I don't know, this would be harder to implement and it's not super necessary, but it would be cool.

Last edited by Coconut Fruit (2019-10-30 18:36:46)

jcwilk · 2019-10-30 18:40:03

jasonrohrer wrote:

jcwilk, the existing algorithm is indeed modeled very closely on Chess Elo rankings.
How much your score jumps after a game of Chess depends on how surprising the result is. If a grandmaster is beaten by the novice, that's a very surprising result, and the novice's score jumps a lot, and the grandmaster's score drops a lot. However, if one grandmaster beats another, that's not a huge surprise, so their scores barely change. The winner goes up a tiny bit, an the loser goes down a tiny bit. Elo has the nice property of converging over time in such a way that it can be used to estimate chances of winning against a given opponent. If we have equal Elo, we are expected to beat each other 50% of the time. A grandmaster (2800) is expected to beat a novice (700) 99.9994% of the time.

How much your score jumps after an offspring dies in OHOL depends on how surprising the result is. Given that your OHOL genetic score is a moving estimate of the average lifespan of your offspring so far, if an offspring exceeds that averge, your score goes up. If an offspring dies much younger than that average, your score goes down.
The top players in OHOL with high scores are like grandmasters in chess. They have nowhere to go but down. If your Elo is 2800, there's no one you can play that will make your ranking go up very much, and a huge number of people you can play (everyone) who will make your ranking go down if you lose to them.
So if your genetic score is 55, that means that on average, you and your offspring have been living 55 years. Living to 60 isn't that surprising, so your score only goes up a little when an offspring lives to 60. But dying at 0 is a HUGE surprise, so your score drops a lot when that happens.

Your example is a bit misleading though, because you're not matching off against other players you're just competing against yourself and your past performance, so it would be like newbie vs newbie or gm vs gm in your chess example. A gm losing to a gm is not a surprising result, it's just fractional ebb and flow of the ones digit, it doesn't throw them careening down to return to the midlevel range with their gm rating long in the past from one reasonable mishap.

Also your experiment seems a little unrealistic, it appears to converge on a certain number but that seems to only be because it's a perfectly cyclical pattern. I bet if you made it 10% failure rather than 9 successes and 1 failure then it would be jumping all over after the occasional 2 or 3 in a row, when in reality even the best player can suffer through having a few babies die due to circumstances outside their control. The big question is whether someone who has a 90% success rate should be consistently rated higher than someone who has 80 or even 70% success rate. With the described algorithm they probably spend more time rated higher than not, but I'd expect them to frequently not sit at #1 despite their superior performance. But maybe that's desired, we seem to be focusing too much on the math and not enough on the ideal behavior so it's difficult to discern what's better or worse.

In another thread I mentioned the possibility of context which is relevant to your example. What if you get born to a really low scoring mother (or even a long line of them, implying a shitty or nonexistent village) but end up being a super achiever as a mother yourself, that would be the equivalent to a newbie getting one over on a gm and deserving a huge boost.

Kaveh · 2019-10-30 18:43:48

Thaulos wrote:

I would guess a quick way of making sure variance isn't too bad but while still providing a quick "startup" would be to somehow factor in the number of lives you previously played.
So maybe first life it would be a big jump, second a big but smaller one, etc. Until you cap variance at the average of your past 10 lives (as an example).
For the score to mean anything it should be hard to drop and to climb. One should only really have high scores if they are consistent across their lives.

I do like the idea of having scores be more stable overtime, so it's less up to luck to be near the top of the leaderboard. An average score for your past 10 lives might indeed work really well, so /die'ing still affects it a lot but only extreme lives will REALLY alter the score.

Thing is, how do you calculate that score?
Is it like a 'grade' out of 60 (to keep the current number)? Does getting to age 60 and all your kids surviving to 60 as well mean you get full points and /die = 0? Then what does a score of 30 mean? Also: what % of the score is just YOUR life and not your kids'? Don't want to incentivice things like avoiding kids because you can always get perfect marks on just your own life without any kids...

Or maybe you start at 30 and early deaths are - and surviving is +, same as now. But if that's the case, does having 8 surviving kids w/ another 2 dying mean you'll get a higher score than having only 2 kids and 0 deaths? Should it be a % of how many of your kids survive? Then what about your own life?

What would be a good way to calculate scores to even get an average from to begin with?

jcwilk · 2019-10-30 18:49:18

Kaveh wrote:

Thaulos wrote:
I would guess a quick way of making sure variance isn't too bad but while still providing a quick "startup" would be to somehow factor in the number of lives you previously played.
So maybe first life it would be a big jump, second a big but smaller one, etc. Until you cap variance at the average of your past 10 lives (as an example).
For the score to mean anything it should be hard to drop and to climb. One should only really have high scores if they are consistent across their lives.
I do like the idea of having scores be more stable overtime, so it's less up to luck to be near the top of the leaderboard. An average score for your past 10 lives might indeed work really well, so /die'ing still affects it a lot but only extreme lives will REALLY alter the score.
Thing is, how do you calculate that score?
Is it like a 'grade' out of 60 (to keep the current number)? Does getting to age 60 and all your kids surviving to 60 as well mean you get full points and /die = 0? Then what does a score of 30 mean? Also: what % of the score is just YOUR life and not your kids'? Don't want to incentivice things like avoiding kids because you can always get perfect marks on just your own life without any kids...
Or maybe you start at 30 and early deaths are - and surviving is +, same as now. But if that's the case, does having 8 surviving kids w/ another 2 dying mean you'll get a higher score than having only 2 kids and 0 deaths? Should it be a % of how many of your kids survive? Then what about your own life?
What would be a good way to calculate scores to even get an average from to begin with?

yeah chess elo traditionally (but iirc not necessarily) does this too, there's an initially faster curve that settles down as you get more games in. Sounds like jason's implementation does this to a degree but dropping 10% from one death 10k births in seems too high still, unless im reading the numbers wrong.

jasonrohrer · 2019-10-30 18:49:48

JC, I will re-run the experiment with true randomness instead of cyclical pattern.

Here's the results after 50,000 babies, with 1/10 babies dying at 0 and 9/10 babies dying at 60, ordered randomly.

Lowest score seen = 31.160344, highest = 59.999215
Overall Ave = 54.108404

jcwilk · 2019-10-30 19:08:36

jasonrohrer wrote:

JC, I will re-run the experiment with true randomness instead of cyclical pattern.

Wish I wasn't moving today, but if I get some time I'll try to help out with some sims too. My favorite tools for these things are jsfiddle.net and desmos.com for code and math respectively since they're both permalinkable

Think you could shed a little light on some of the desired attributes of the ideal algorithm? Or even if there were a few key contrived situations with desired consequences we could focus on. The current usage of it is just for tool slots or for eve selection also? Is eve selection something that should be really strict about only the best available players getting a chance to do or would some unpredictability about who gets chosen be desired?

For these kinds of things I usually like to have a core algorithm that's as deterministic as possible, and then layer on top of that any randomness I want so that I can scale the randomness easily and still always have the "truth" to refer to, in this case that might mean trying to differentiate the 95% people from the 90% people in a way that doesn't fluctuate unless their actual ability changes, ie, unless they start consistently doing worse after a long period of consistent success. That might not be possible but it's just an example of an ideal to work towards, or maybe that's not the ideal you had in mind at all. Hard to pick the right math without knowing that.

jcwilk · 2019-10-30 19:10:31

jasonrohrer wrote:

JC, I will re-run the experiment with true randomness instead of cyclical pattern.
Here's the results after 50,000 babies, with 1/10 babies dying at 0 and 9/10 babies dying at 60, ordered randomly.
Lowest score seen = 31.160344, highest = 59.999215
Overall Ave = 54.108404

Yeah but you're only looking at the average, that's not what I was talking about... I mentioned concern about whether someone who was consistently better than someone else would consistently be rated higher than they are, or whether their rating would wildly fluctuate. Like look at the numbers in the last 100 or so babies in that 50k trial, my guess is that it dips quite low in some points.

Edit: oops, missed the lowest,highest. Yeah I mean rating of 31 for someone who on average has 54 is not great right? 54 is quite high out of 60.

Edit edit: Depends on where it happened, but yeah i feel like the first 100 births and last 100 births of a long simulation are going to be the interesting bits. Maybe it would be worth defining how long it should take to get to 55, how many babies in a row need to die in order to drop someone below 50 after being a long time 55er, whether losing 10 babies in one life is as bad as losing 1 baby each life across 10 lives, etc

Last edited by jcwilk (2019-10-30 19:17:22)

jasonrohrer · 2019-10-30 19:15:20

So here's the thing, though... we say, "Do we want the good player to always have a higher ranking than the bad player?" Yes, it seems like we do. Having only 1/10 of your BB die will eventually get you to the very very top of the leaderboard with a score of 59.999. Of course, it will also take you down to 31.16 sometimes too.

Lets look at the ranges for other BB death rates, though...

Baby death rate 0.1, Lowest score seen = 28.411339, highest = 59.998843
                     Overall Ave = 54.068327

Baby death rate 0.2, Lowest score seen = 20.184815, highest = 59.861929
                     Overall Ave = 47.764049

Baby death rate 0.3, Lowest score seen = 14.330883, highest = 59.788322
                     Overall Ave = 41.950563

Baby death rate 0.4, Lowest score seen = 10.048529, highest = 56.575497
                     Overall Ave = 36.128236

Baby death rate 0.5, Lowest score seen = 6.213488, highest = 53.046858
                     Overall Ave = 30.123685

Baby death rate 0.6, Lowest score seen = 3.630818, highest = 51.230885
                     Overall Ave = 23.924169

Baby death rate 0.7, Lowest score seen = 0.504421, highest = 47.674427
                     Overall Ave = 17.880723

Baby death rate 0.8, Lowest score seen = 0.185903, highest = 41.280818
                     Overall Ave = 12.095610

Baby death rate 0.9, Lowest score seen = 0.002793, highest = 30.551247
                     Overall Ave = 6.038323

These are all over runs for 50,000 babies, after an initial warm up run of 50,000 babies to let the scores settle.

The averages are perfect, so that's good.

But the variance is huge. It's pretty bad that the 0.1, 0.2, and 0.3 players all make it to the very top of the leaderboard.

I think that we might just need to adjust the K factor, which is currently 10, and see how that changes things. Bigger K means less jumpyness. The formula is:

new_score = old_score + ( ( lifespan - old_score ) ^ R ) / K

This is the Elo update formula, with the extra R factor. R is currently 1, so has no effect (also, there's a tweak in the code to preserve sign if R is 2).

Thaulos · 2019-10-30 19:20:56

Ideally the range of the score would be lower. Can't you add in some sort of average?

Last edited by Thaulos (2019-10-30 19:21:14)

jasonrohrer · 2019-10-30 19:22:03

Here's K = 20:

Baby death rate 0.1, Lowest score seen = 40.673095, highest = 59.918052
                     Overall Ave = 53.945883

Baby death rate 0.2, Lowest score seen = 32.225654, highest = 59.030679
                     Overall Ave = 47.843927

Baby death rate 0.3, Lowest score seen = 21.538847, highest = 55.594381
                     Overall Ave = 42.144271

Baby death rate 0.4, Lowest score seen = 18.403748, highest = 51.936603
                     Overall Ave = 36.116350

Baby death rate 0.5, Lowest score seen = 12.244200, highest = 48.589418
                     Overall Ave = 30.011191

Baby death rate 0.6, Lowest score seen = 8.143825, highest = 42.203674
                     Overall Ave = 23.957806

Baby death rate 0.7, Lowest score seen = 4.865320, highest = 36.375087
                     Overall Ave = 17.990757

Baby death rate 0.8, Lowest score seen = 1.349298, highest = 29.174280
                     Overall Ave = 11.804335

Baby death rate 0.9, Lowest score seen = 0.104732, highest = 21.201887
                     Overall Ave = 5.803208

jcwilk · 2019-10-30 19:22:56

jasonrohrer wrote:

So here's the thing, though... we say, "Do we want the good player to always have a higher ranking than the bad player?" Yes, it seems like we do. Having only 1/10 of your BB die will eventually get you to the very very top of the leaderboard with a score of 59.999. Of course, it will also take you down to 31.16 sometimes too.
Lets look at the ranges for other BB death rates, though...
Baby death rate 0.1, Lowest score seen = 28.411339, highest = 59.998843
                     Overall Ave = 54.068327

Baby death rate 0.2, Lowest score seen = 20.184815, highest = 59.861929
                     Overall Ave = 47.764049

Baby death rate 0.3, Lowest score seen = 14.330883, highest = 59.788322
                     Overall Ave = 41.950563

Baby death rate 0.4, Lowest score seen = 10.048529, highest = 56.575497
                     Overall Ave = 36.128236

Baby death rate 0.5, Lowest score seen = 6.213488, highest = 53.046858
                     Overall Ave = 30.123685

Baby death rate 0.6, Lowest score seen = 3.630818, highest = 51.230885
                     Overall Ave = 23.924169

Baby death rate 0.7, Lowest score seen = 0.504421, highest = 47.674427
                     Overall Ave = 17.880723

Baby death rate 0.8, Lowest score seen = 0.185903, highest = 41.280818
                     Overall Ave = 12.095610

Baby death rate 0.9, Lowest score seen = 0.002793, highest = 30.551247
                     Overall Ave = 6.038323
These are all over runs for 50,000 babies, after an initial warm up run of 50,000 babies to let the scores settle.

The 0.5 run seems undesirable, they shouldn't get a rating of 53 when their average is 30... But the highest number across such a wide range of randomness can be deceiving, maybe the second run which is the measured run should only be 100, maybe do a few trials of each and average the highest and lowest values? I'm a bit out of my element here with stats, but having "good player should always be higher than worse player" as a starting point is something. Maybe we can keep working towards more goals like that and then rally the coders/mathies in the community to come up with an algorithm that fits

jasonrohrer · 2019-10-30 19:23:54

Here's K = 40:

Baby death rate 0.1, Lowest score seen = 45.393603, highest = 59.010466
                     Overall Ave = 54.074477

Baby death rate 0.2, Lowest score seen = 35.006117, highest = 56.155618
                     Overall Ave = 47.975000

Baby death rate 0.3, Lowest score seen = 29.719307, highest = 52.229610
                     Overall Ave = 41.833456

Baby death rate 0.4, Lowest score seen = 21.244996, highest = 47.863943
                     Overall Ave = 36.028147

Baby death rate 0.5, Lowest score seen = 15.803323, highest = 42.724206
                     Overall Ave = 30.242026

Baby death rate 0.6, Lowest score seen = 11.069928, highest = 36.938155
                     Overall Ave = 23.983215

Baby death rate 0.7, Lowest score seen = 6.946684, highest = 30.186798
                     Overall Ave = 17.952708

Baby death rate 0.8, Lowest score seen = 3.594163, highest = 27.430505
                     Overall Ave = 12.010568

Baby death rate 0.9, Lowest score seen = 0.593459, highest = 15.735030
                     Overall Ave = 6.094836

jcwilk · 2019-10-30 19:24:26

jasonrohrer wrote:

The averages are perfect, so that's good.
But the variance is huge. It's pretty bad that the 0.1, 0.2, and 0.3 players all make it to the very top of the leaderboard.
I think that we might just need to adjust the K factor, which is currently 10, and see how that changes things. Bigger K means less jumpyness. The formula is:
new_score = old_score + ( ( lifespan - old_score ) ^ R ) / K
This is the Elo update formula, with the extra R factor. R is currently 1, so has no effect (also, there's a tweak in the code to preserve sign if R is 2).

Agree, sounds good. Bumping up the K value and then taking a look at the first few dozen could be useful too (edit: ie, before they've settled since that's the biggest downside of a high K value, unless you want to gradually raise the K value the more they play which is I think how chess usually does it)

Last edited by jcwilk (2019-10-30 19:25:43)

jasonrohrer · 2019-10-30 19:25:06

Here's K=80

Baby death rate 0.1, Lowest score seen = 46.033451, highest = 58.470665
                     Overall Ave = 53.927025

Baby death rate 0.2, Lowest score seen = 41.245160, highest = 54.579930
                     Overall Ave = 48.115154

Baby death rate 0.3, Lowest score seen = 32.956241, highest = 50.164080
                     Overall Ave = 41.848375

Baby death rate 0.4, Lowest score seen = 27.565044, highest = 44.636237
                     Overall Ave = 35.949929

Baby death rate 0.5, Lowest score seen = 20.806898, highest = 38.819771
                     Overall Ave = 30.091764

Baby death rate 0.6, Lowest score seen = 15.473006, highest = 32.202993
                     Overall Ave = 23.796076

Baby death rate 0.7, Lowest score seen = 10.184094, highest = 26.023644
                     Overall Ave = 18.045427

Baby death rate 0.8, Lowest score seen = 6.133832, highest = 19.024386
                     Overall Ave = 12.076141

Baby death rate 0.9, Lowest score seen = 1.730073, highest = 11.139416
                     Overall Ave = 6.033944

One Hour One Life Forums

#1 2019-10-30 16:18:26

If genetic fitness actually matters, is life limit still necessary?

#2 2019-10-30 16:31:40

Re: If genetic fitness actually matters, is life limit still necessary?

#3 2019-10-30 16:34:04

Re: If genetic fitness actually matters, is life limit still necessary?

#4 2019-10-30 16:48:16

Re: If genetic fitness actually matters, is life limit still necessary?

#5 2019-10-30 16:59:57

Re: If genetic fitness actually matters, is life limit still necessary?

#6 2019-10-30 17:19:54

Re: If genetic fitness actually matters, is life limit still necessary?

#7 2019-10-30 17:35:12

Re: If genetic fitness actually matters, is life limit still necessary?

#8 2019-10-30 17:47:09

Re: If genetic fitness actually matters, is life limit still necessary?

#9 2019-10-30 17:47:41

Re: If genetic fitness actually matters, is life limit still necessary?

#10 2019-10-30 17:47:49

Re: If genetic fitness actually matters, is life limit still necessary?

#11 2019-10-30 18:17:15

Re: If genetic fitness actually matters, is life limit still necessary?

#12 2019-10-30 18:35:44

Re: If genetic fitness actually matters, is life limit still necessary?

#13 2019-10-30 18:40:03

Re: If genetic fitness actually matters, is life limit still necessary?

#14 2019-10-30 18:43:48

Re: If genetic fitness actually matters, is life limit still necessary?

#15 2019-10-30 18:49:18

Re: If genetic fitness actually matters, is life limit still necessary?

#16 2019-10-30 18:49:48

Re: If genetic fitness actually matters, is life limit still necessary?

#17 2019-10-30 19:08:36

Re: If genetic fitness actually matters, is life limit still necessary?

#18 2019-10-30 19:10:31

Re: If genetic fitness actually matters, is life limit still necessary?

#19 2019-10-30 19:15:20

Re: If genetic fitness actually matters, is life limit still necessary?

#20 2019-10-30 19:20:56

Re: If genetic fitness actually matters, is life limit still necessary?

#21 2019-10-30 19:22:03

Re: If genetic fitness actually matters, is life limit still necessary?

#22 2019-10-30 19:22:56

Re: If genetic fitness actually matters, is life limit still necessary?

#23 2019-10-30 19:23:54

Re: If genetic fitness actually matters, is life limit still necessary?

#24 2019-10-30 19:24:26

Re: If genetic fitness actually matters, is life limit still necessary?

#25 2019-10-30 19:25:06

Re: If genetic fitness actually matters, is life limit still necessary?

Board footer