36 coaches online • Server time: 11:07
Forum Chat
Log in
Recent Forum Topics goto Post DOTP Season 4goto Post Skittles' Centu...goto Post Secret League Americ...
SearchSearch 
Post new topic   Reply to topic
View previous topic Log in to check your private messages View next topic
Macabeo



Joined: Feb 13, 2011

Post   Posted: Jul 19, 2019 - 12:04 Reply with quote Back to top

Wozzaa wrote:
Did anyone see this blog?


This. The first sensible step is participating in https://bot-bowl.com, or at least following it closely.
tussock



Joined: May 29, 2011

Post   Posted: Jul 19, 2019 - 14:34 Reply with quote Back to top

Hmm, depends on movement speed, say all-skink teams, but players can move to any of seven spaces a number of times equal to their MA, times 2^4(DT+BT+Dodge+Leap), plus each step can stop, prone, stun, or off, which is 7*16*5 per step, plus any step can be a block, because you have to check all of them for the optimal blitz even though you'll only pick one.

Blocking is like another 68 possible states with grab/sidestep (up/down/stun/ko+, either or, pushes), oh, plus chainpushes gives us ... 194 more ways for each of 12 of them (pushed or surfed), I think at most push five guys. Oh, maybe each push changes 10 assist spaces, local down changes 7. That's 1+55*7+12*194*10 = 24170 results from a block.

So ... (24170+16*7*5)^8 = 1.4*10^35.


Hmm, only one pickup per turn, and while infinite failed catches are possible, it's only ~120 places the ball might stop, again for the handoff, and 390 for the pass for finishing square, *2 for accurate or not ...

So (2*390+2*120+1) = 1021 ball variations have their own odds from each step, as anyone may have the ball even picking up on a blitz, and handing off or passing is a consideration each player can make and it all changes depending on other results, so multiplies.

1.4*10^38 ways to move players, make blocks, take the blitz, adjust opponents, and move the ball around. The foul could technically clear a space to change odds for any of that, so *3 (space clear or not clear or ends turn).

4.3*10^38 including the foul, after (from Nelphine's post) ~7.5*10^6 possible turn-starting positions is 3.3*10^45 for the last turn in a half.

Each turn can be no more complex than that, considering only itself, thus considering others, you count back up, and add complexities.

So *16! would be the right multiplier, earlier turns are more complex. The first turn considers 16 turns, the next 15 turns, and so on. Then double the total for the second half.

Should be ~1.4*10^59 ways to play any particular game.


Oh, probably *8 for shadowing, which is ~1.1*10^60.

--

And if you score lots? Uh, ... I think that makes it much, much simpler. It should do, duelling 2-turns are obviously many orders simpler in play.

Possibly gutter runners and the MA 10 freaks would tend to increase complexity a bit, which, probably they do, yeah, from the way it feels to try and cover them on pitch.

Things like guard don't strictly increase the game space, they just make each particular step much harder to calculate.

I just don't see it being backgammon. A giant mess of guards in the middle is difficult to understand, but it doesn't change any more than anything else in the upper limit.

10^60 games for 10^50 pairs of teams, is 10^110 games of bloodbowl. That's not bad.

_________________
ImageImage
Endzone



Joined: Apr 01, 2008

Post   Posted: Jul 19, 2019 - 17:09 Reply with quote Back to top

I suspect AI could be developed to be a lot better than all of us but as a niche game the question is if and when anyone will actually bother to do it. I remember 20 years ago everyone saying that a computer would never beat the best chess players because strategy would always beat tactics. But when you can calulate tactics like a computer tactics dictate the strategy. A fun paper for someone to write!
thoralf



Joined: Mar 06, 2008

Post   Posted: Jul 19, 2019 - 19:31 Reply with quote Back to top

Nelphine wrote:
So, at this point, we have (24*10)*10*10*10*10*10*10*10*10*10*10*10*10*10*10*10*(24*10)*10*10*10*10*10*10*10*10*10*10*10*10*10*10*10 choices of players. Or 496*10^32.


We could push this argument further by adding SL teams, Stunty League, or just go with the Cartesian product of all possible statistics with all possible skills and beyond.

Let's simplify. Each coach can still only field 11-12 players. The only relevant characteristic is the MA. Everything else only impacts the odds of success of an action. So assume they all have 10MA and always move freely on the whole 26x15 field. That's 390 possible moves per turn per player, including doing nothing and allowing the possibility that all players stand on the same square. Assume also 20-turns games with no removal.

That gives us 22 players ^ 390 possible moves ^ 40 turns. Calculator.net gives me

62,204,360,041,003,679,337,192,591,473,285,100,401,
621,229,306,513,879,331,423,044,053,960,226,676,723,
006,562,030,192,940,286,485,959,282,734,790,063,206,
421,952,096,734,243,048,199,611,217,824,725,748,023,
008,471,914,972,935,321,953,469,450,053,606,570,589,
556,544,475,717,572,115,403,458,811,296,233,654,538,
739,269,609,327,528,456,588,158,346,156,905,929,295,
733,213,031,428,680,848,482,020,448,064,153,501,238,
335,866,580,064,735,208,727,796,759,331,616,997,197,
246,642,635,509,040,052,398,558,229,712,145,065,844,
545,361,707,580,942,092,615,230,612,038,454,133,133,
093,434,635,803,900,967,123,524,713,255,878,541,655,
412,637,418,254,996,192,329,423,622,478,025,727,363,
636,191,850,119,541,742,900,757,946,503,614,715,044,
191,868,052,410,397,416,662,346,788,616,888,665,499,
869,666,082,470,512,443,633,377,941,800,839,419,646,
892,128,818,227,537,929,485,524,584,086,824,546,639,
184,541,835,363,507,323,714,388,622,277,229,973,416,
623,461,506,092,216,880,416,870,920,635,548,573,368,
264,858,381,623,437,569,400,967,400,326,906,584,136,
513,704,228,685,166,707,728,660,329,588,502,689,124,
101,943,987,004,957,481,560,209,269,598,201,669,676,
622,083,595,828,259,870,646,653,387,559,861,749,660,
224,538,280,308,574,560,818,475,198,904,917,973,782,
829,841,066,451,529,596,596,164,726,718,248,086,398,
802,086,175,638,798,906,437,161,194,946,577,472,488,
148,400,802,547,054,863,436,820,875,270,423,114,782,
980,737,877,649,701,879,833,983,865,094,433,404,127,
563,051,265,237,030,801,130,109,715,854,287,787,248,
338,574,275,188,098,449,259,530,904,863,306,496,327,
199,996,997,558,509,409,363,065,342,532,571,869,826,
493,971,899,071,987,419,110,881,985,545,785,671,704,
200,199,191,755,634,032,099,630,060,378,924,951,785,
975,269,805,688,445,124,798,990,931,439,866,019,726,
782,241,220,214,397,955,299,478,374,078,388,495,717,
613,763,825,699,289,749,893,135,800,117,322,489,555,
747,683,188,787,403,262,445,154,020,276,416,537,740,
680,092,705,046,563,669,323,592,513,421,897,900,211,
052,021,815,437,231,268,757,236,507,873,644,189,156,
029,664,875,217,131,707,631,655,177,407,391,792,973,
928,935,406,742,379,197,623,965,917,819,347,399,225,
359,519,409,188,277,793,935,711,577,305,581,806,607,
342,414,110,457,598,006,225,758,744,832,132,894,673,
211,463,285,988,521,905,012,084,319,248,260,478,415,
126,880,371,831,383,329,791,355,165,593,631,148,620,
420,648,378,920,874,007,703,975,192,434,912,403,547,
234,799,879,772,737,410,317,734,498,750,922,763,758,
884,986,252,402,249,243,068,151,379,243,124,128,222,
391,977,181,852,890,291,531,594,314,881,566,054,668,
889,963,143,857,840,538,016,030,528,451,876,698,513,
121,980,837,822,789,543,772,537,646,533,973,277,989,
567,309,106,784,906,337,007,649,829,996,915,879,923,
955,999,010,093,800,270,951,914,644,826,001,280,778,
395,679,941,647,408,093,260,099,892,104,942,923,837,
699,673,436,132,514,271,983,246,433,157,624,136,896,
426,914,754,395,652,522,846,767,931,912,227,431,502,
802,608,361,575,241,033,520,652,190,137,056,883,964,
384,040,052,654,375,356,628,549,791,308,878,315,451,
046,418,263,784,363,462,696,428,516,480,846,601,985,
888,343,578,012,385,906,247,869,458,774,897,990,975,
746,836,856,710,446,549,241,342,013,854,508,941,506,
590,464,173,487,415,483,965,349,254,028,336,962,910,
243,433,252,845,502,293,882,431,317,782,446,137,172,
499,471,962,498,876,350,694,561,427,884,738,274,749,
121,116,738,543,022,079,021,429,525,416,227,521,784,
997,826,084,209,189,243,422,018,337,504,529,658,847,
848,498,860,579,857,554,869,211,539,411,006,094,935,
812,204,520,350,918,402,175,553,148,098,223,229,799,
013,792,814,020,761,796,500,241,040,107,147,386,554,
562,436,342,085,720,323,148,417,401,604,859,688,231,
099,795,832,459,362,141,274,017,786,604,260,857,437,
402,293,327,089,419,747,588,766,142,595,430,642,434,
298,802,439,503,253,705,617,558,821,267,372,843,014,
410,956,484,168,691,472,640,235,939,527,016,075,624,
167,954,792,099,761,694,064,523,440,104,813,667,660,
647,243,691,069,386,088,146,860,273,446,570,140,670,
463,784,065,047,313,277,694,993,598,732,665,065,482,
345,604,551,647,896,477,931,826,999,107,065,845,732,
961,776,130,156,615,070,738,665,647,130,173,364,095,
054,662,900,598,468,998,468,087,458,430,659,754,382,
683,398,640,055,335,761,629,541,288,601,228,662,780,
685,080,178,182,401,003,001,392,726,693,226,749,288,
037,726,292,065,409,438,678,723,777,434,818,794,727,
957,063,941,044,980,518,569,236,193,816,479,830,232,
320,013,280,433,569,116,664,415,705,030,367,238,309,
715,528,557,637,279,861,862,558,023,359,673,601,089,
147,282,139,619,854,456,537,134,287,255,177,630,280,
587,151,502,262,188,363,144,574,040,428,545,990,490,
496,837,328,668,815,129,709,409,100,145,438,030,416,
461,758,454,252,387,915,436,930,929,549,597,830,445,
700,259,071,102,730,507,559,580,892,445,934,067,430,
868,282,183,135,763,580,607,741,693,187,480,109,317,
928,792,819,351,049,736,606,802,317,755,792,332,153,
460,305,014,514,532,932,286,008,387,124,026,011,943,
529,820,619,046,589,804,048,119,489,715,849,369,072,
956,671,044,223,818,905,152,136,831,583,086,275,969,
881,565,521,917,376,160,440,848,502,404,604,600,616,
683,816,316,238,931,602,402,322,555,616,596,405,081,
198,381,550,930,331,670,269,036,486,242,570,982,701,
289,969,342,435,861,047,137,518,874,131,151,697,049,
907,144,653,229,231,456,120,099,162,578,175,453,278,
075,619,588,297,233,064,284,660,788,485,544,227,271,
810,430,022,667,908,660,598,971,955,112,709,185,792,
936,426,778,428,711,147,487,396,442,135,136,599,555,
072,012,866,723,615,083,526,941,217,505,018,747,807,
233,388,585,598,335,284,636,196,405,463,750,899,350,
385,817,008,988,002,226,109,029,022,885,838,698,568,
337,511,366,748,916,023,343,992,539,898,077,840,421,
943,326,382,788,558,861,683,433,724,394,398,376,774,
663,078,370,697,637,220,227,157,345,521,006,109,992,
656,147,001,935,526,367,060,532,969,379,015,116,926,
150,624,267,313,783,229,880,296,990,748,823,109,743,
172,629,672,582,583,124,516,471,903,998,024,604,010,
242,163,493,502,850,122,400,501,333,541,276,588,932,
939,617,079,469,584,921,534,101,812,058,931,805,885,
752,414,554,640,591,616,896,781,420,811,865,863,154,
253,700,068,004,091,679,615,756,426,124,408,890,666,
130,459,572,668,294,839,848,334,958,618,799,919,369,
893,328,335,205,105,555,688,821,733,679,561,838,895,
745,191,445,489,940,544,125,360,341,551,465,185,608,
304,880,622,739,272,759,088,714,640,251,591,087,259,
940,528,975,741,911,318,209,904,270,024,959,742,394,
188,875,371,879,041,447,120,857,771,200,339,489,703,
481,358,602,205,962,994,083,311,130,070,618,147,302,
456,741,175,034,108,311,799,457,345,057,707,896,135,
218,925,969,654,368,251,366,231,749,672,325,745,004,
103,109,550,994,407,422,182,977,116,477,803,939,270,
511,798,631,864,308,259,280,411,492,185,492,846,340,
114,640,525,773,785,711,107,442,594,423,353,533,635,
754,432,554,660,925,040,693,402,820,840,585,036,685,
983,941,924,798,821,148,553,217,898,763,191,044,749,
930,442,589,660,221,505,143,269,182,432,111,137,545,
564,315,472,716,004,281,043,002,334,615,668,679,089,
950,513,474,531,879,224,846,024,689,859,277,431,032,
513,914,467,925,743,805,700,899,426,904,228,922,584,
287,411,190,395,907,017,625,728,415,983,815,520,106,
596,093,094,397,869,807,813,519,414,780,917,096,787,
368,393,360,366,733,997,196,958,573,411,521,694,228,
937,500,038,896,849,020,101,070,876,618,867,537,439,
928,312,740,669,086,223,497,167,886,317,680,622,797,
667,450,747,741,193,522,532,739,240,953,509,187,081,
821,192,855,093,210,448,221,611,280,744,238,581,536,
399,143,311,155,183,779,992,847,535,562,096,676,722,
222,757,493,770,961,104,970,127,261,012,172,386,587,
100,268,127,399,308,759,138,815,292,237,751,408,383,
716,621,079,580,307,934,128,918,774,953,857,902,318,
051,471,982,618,874,070,667,027,788,332,588,569,798,
006,666,091,785,879,682,630,400,922,500,438,489,278,
512,544,263,875,110,363,247,235,772,147,994,276,923,
249,548,442,712,007,033,408,133,242,912,861,681,405,
532,742,753,377,579,419,889,740,070,169,304,397,627,
297,068,943,975,097,393,490,017,668,916,962,153,091,
483,958,061,801,645,056,215,231,279,045,872,705,139,
889,687,584,492,367,500,907,082,729,641,210,681,428,
856,302,533,875,951,251,092,212,735,159,718,240,259,
912,221,065,697,908,860,047,544,986,558,555,337,563,
452,061,488,243,135,806,070,106,916,676,055,058,295,
626,299,070,603,097,387,176,845,939,694,869,718,032,
562,318,544,731,810,466,294,647,571,504,771,168,311,
841,231,894,606,621,110,088,297,520,711,997,031,219,
255,214,436,808,020,803,714,212,862,240,715,374,448,
259,599,841,992,231,490,851,154,742,938,771,056,235,
717,002,314,507,555,679,732,819,650,594,275,694,324,
923,218,023,960,132,466,037,240,715,550,038,586,757,
732,903,139,878,996,654,395,567,074,345,809,818,026,
974,727,801,309,193,939,455,772,837,805,223,661,201,
763,098,722,951,308,823,965,083,254,009,615,364,628,
438,422,341,384,532,288,677,071,896,724,359,439,645,
875,822,395,709,493,642,396,298,023,703,790,345,164,
225,293,812,459,544,416,915,858,861,203,452,310,855,
737,959,111,364,199,775,602,940,598,508,364,082,137,
690,847,273,650,183,796,880,390,795,699,800,804,760,
796,896,133,990,016,998,422,665,631,710,452,125,733,
770,352,782,025,504,665,052,683,184,130,863,098,773,
943,619,878,214,748,650,035,814,229,415,619,731,128,
199,513,275,362,486,902,451,617,966,587,634,384,819,
341,693,372,101,367,213,384,075,841,222,618,910,991,
486,466,251,622,096,626,386,567,051,758,559,200,722,
057,087,605,287,741,454,410,251,966,773,701,398,380,
920,132,485,004,929,598,953,322,799,226,578,714,646,
816,224,709,045,911,282,255,328,230,525,790,622,577,
990,453,693,535,442,394,998,899,988,092,110,778,915,
370,781,839,521,722,825,669,132,611,359,865,088,872,
662,098,653,797,471,887,493,028,176,575,726,804,929,
578,002,881,703,535,731,589,668,501,962,596,878,263,
644,051,467,079,417,461,761,392,969,929,918,533,528,
445,853,806,921,260,176,056,725,310,773,250,038,512,
808,711,710,593,513,789,855,088,567,137,113,547,861,
784,269,445,777,763,763,386,043,428,596,875,799,899,
265,948,740,937,001,693,062,788,867,897,676,052,238,
348,830,768,294,210,558,304,587,491,362,570,392,328,
082,107,679,544,155,715,957,850,861,875,889,268,467,
007,075,670,347,012,221,811,193,507,748,560,512,037,
893,864,679,403,810,833,300,716,293,676,219,314,927,
305,202,101,963,656,550,590,073,335,419,457,833,309,
863,080,059,110,715,175,105,267,066,302,224,075,516,
328,490,430,028,721,135,114,077,270,996,085,016,478,
558,756,527,662,246,545,491,999,842,455,466,172,119,
783,520,267,732,396,884,992,755,831,259,235,616,501,
470,160,747,107,349,679,693,052,980,897,778,823,864,
870,370,005,877,957,812,474,375,074,238,377,602,530,
616,431,822,039,798,645,296,087,462,144,562,454,122,

et cetera. That's just about half the number.

ADD. See Note.

***

Most perceptual tasks are more complex than that.

NB. This number was slightly exagerrated.

_________________
There is always Sneaky Git.


Last edited by thoralf on Jul 20, 2019 - 04:39; edited 1 time in total
JellyBelly



Joined: Jul 08, 2009

Post   Posted: Jul 19, 2019 - 19:48 Reply with quote Back to top

You would need quite a lot of RAM for that Wink

Are there even that many atoms in the known universe?

_________________
"Opinions are like arseholes, everybody's got them and they all stink." - The protagonist, Fallout 2

"Go for the eyes, Boo! Go for the eyes!!" Razz
thoralf



Joined: Mar 06, 2008

Post   Posted: Jul 19, 2019 - 20:15 Reply with quote Back to top

JellyBelly wrote:
Are there even that many atoms in the known universe?


There are aleady more Chess games than there are atoms in the known universe, so this number is really BIG.

A good video on Shannon Number:

https://www.youtube.com/watch?v=Km024eldY1A

The search made me find this very elegant page on various notations for truly big numbers:

https://www-users.cs.york.ac.uk/~susan/cyc/b/big.htm

***

These numbers are quite vast, but consider a simple 50-pieces jigzaw puzzle. Even if you're a perfect player, there are something like 50^50 different "routes" to solve it. [Add: it's a really rough estimate.] To get a tower of exponentials, all we need is to add rotations, more realistic configurations, takebacks and so on. Yet kids do that without thinking too much.

Now, consider a vacuum bot. The only way the bot can go on and about your house is by "abstracting away" most of the information that it could in principle process. Imagine all the possible house configurations for which it's designed. A LOT bigger than a 26 x 15 game!

_________________
There is always Sneaky Git.


Last edited by thoralf on Jul 20, 2019 - 04:40; edited 2 times in total
Nelphine



Joined: Apr 01, 2011

Post   Posted: Jul 19, 2019 - 20:53 Reply with quote Back to top

I'm.not sure I follow you. We have 6*10^56 initial start positions.

Now, on a given turn, for an upper bound, each player could move up to 13 times (MA 10 + 3 gfi's via sprint).

Each move has up to 24 possible choices (due to leap).

Each of those can be impacted by the position, stats, and skills of all the other players, as well as the players own position stats and skills.

Therefore, the possible moves for any given player are 6*10^56 * 13 * 24.

Then, there are up 12 players that can move on any given turn.

The order of players absolutely matters, so there are 12! * 6*10^56*13*24 possible moves for a given turn.

Each move is then split into nothing, cannot act, move, blitz, pass, hand off. (There are other action types, but they should be subsumed by these given were talking about an upper bound).

Therefore, for any given turn, there are 6*12!*6^*10^56 possible actions, *6*10^56 possible start states, = 3.2*10^125 possible initial game states, where initial game state includes possible moves from that game state.

This is actually lower than he 10^200 I guessed at last time, but it's still fairly complex.
Nelphine



Joined: Apr 01, 2011

Post   Posted: Jul 19, 2019 - 20:59 Reply with quote Back to top

I would also assume 30 turn games not 20 turns when trying to calculate game size, due to overtime. You'd further ant to account for current score for game size, which I haven't yet. You'd also want to calculate for opponents moves, so a total of 60 turns. And you'd want to account for team manager priorities, for things like 'is this part of a tournament? Is it a one loss ko tournament? Does spp accumulation matter? Does saving a certain player matter?' etc.
Which in itself is a complex game, further increasing both the possible moves on a given turn beyond the ^125 I currently have, but also the game size as well.
thoralf



Joined: Mar 06, 2008

Post   Posted: Jul 19, 2019 - 23:32 Reply with quote Back to top

Nelphine wrote:
I'm.not sure I follow you. We have 6*10^56 initial start positions.

I was trying to estimate an upper bound by abstracting away everything except movement. I also assumed omnipotence, i.e. a player can move from any one square to any other one. Like a Chess program would calculate all the Queen's moves in a virtual board and then substract what lies outside the "real" board, but without the substracting part.

This way there's no need need to distinguish a central square from a corner square. An MA 6 guy in a corner can reach 100 squares with gfis and Sprint (a 10 x 10 zone) whereas he can reach 285 squares in the middle of the field (a 19 x 15 zone).

***

No big number should really increase complexity, as we have results on N x N Chess. Once we accept that the game is complex, everything that lies outside the evaluation function's horizon becomes moot. Only how the function gets perfected matters. Alpha will beat us either it's Chess, Go, or BB.

So I told myself - what the heck, let's show this terrific calculator and this really big number.

I don't know any work that ever succeeded in doing for Chess what you're trying to do with your BB calculations. I haven't checked them, but if your numbers are correct, you got yourself a paper.

At the end of the Numberphile video, there was a mention of Hardy's number, i.e. 10^10^50. That's way bigger than Shannon's estimate of 10^43. Shannon's point was to suggest that no exhaustive method would work. He was right.

_________________
There is always Sneaky Git.


Last edited by thoralf on Jul 19, 2019 - 23:52; edited 2 times in total
thoralf



Joined: Mar 06, 2008

Post   Posted: Jul 19, 2019 - 23:48 Reply with quote Back to top

Some loose ends:

Nelphine wrote:
I would also assume 30 turn games not 20 turns when trying to calculate game size, due to overtime.

Correct. Hence why I used 40 in my estimate: (8+8+4) x 2.

Nelphine wrote:
You'd further ant to account for current score for game size, which I haven't yet.

That one will be tough. You'll need a program to "witness" these scores. We know exhaustive methods won't do. Perhaps a real programmer could help you here.

Nelphine wrote:
You'd also want to calculate for opponents moves, so a total of 60 turns.

There can't be more than 48 turns under the current rules.

Nelphine wrote:
And you'd want to account for team manager priorities, for things like 'is this part of a tournament? Is it a one loss ko tournament? Does spp accumulation matter? Does saving a certain player matter?' etc.

Chess engines never entertain meta-game considerations. I suspect they can be included in the risk vs reward ratio I alluded earlier. If the AI already wins 3-0, it should not leap into cages just for the fun of it.

There may be a parameter that could do the trick:

Quote:
The Contempt Factor reflects the estimated superiority/inferiority of the program over its opponent. The Contempt factor is assigned as draw score to avoid (early) draws against apparently weaker opponents, or to prefer draws versus stronger opponents otherwise.


https://www.chessprogramming.org/Contempt_Factor

_________________
There is always Sneaky Git.
Nelphine



Joined: Apr 01, 2011

Post   Posted: Jul 20, 2019 - 00:19 Reply with quote Back to top

Why would you need a 'witness' for the scores? you just make the difference between your score and your opponent's score another factor in the game state complexity, similar to weather. Probably reasonable to say tied, +-1, +-2, +-3 (or more), so 7 states. Multiply my number by 7.

However, you'd also need to keep track of how many turns are left in the drive, which can be up to 9, so another 9 states. Multiply my number by another 9.

With riots, you can absolutely have more than 24 turns in a game. Strictly, I'm not aware of any rule that prevents there being infinite turns if you have duelling one turners, but for the purposes of the discussion, assuming no more than 3 riots is reasonably safe. Might want to actually assume 5 but *shrug* it really doesn't change much anyway.

Meta-game would absolutely need to be a seperate game. However, once the meta-game ai has also be designed, it should be designed to provide outputs. These outputs would be the 'moves' for the meta-game game, and they would translate to 'goals' for the bloodbowl game.

In other words, the meta-game AI would do all its stuff, and determine that in the next game, the best 'move' in order to win the overall tournament or season or maximize CR (it would need a defined goal) , might be to win the next game AND get at least 1 SPP on the Saurus who has 5 SPP.

Then this 'move' would be fed into the bloodbowl game AI, who would then consider optimal moves, but not just on winning, but ALSO on getting 1 SPP.

This would mean that the 'win' state for the bloodbowl AI would be FAR more complex than almost any game so far considered on the various webpages linked, because the win state might NOT be just 'win the game'.

Therefore the optimal route to 'win' would be much more complex, as it would have variable conditions.

In certain cases the 'win' goal provided by the meta-game AI, might NOT actually include winning the next game. For instance, if the AI is playing in a 9 game season, and it has already won 8 games, and the players with most games make it to the play offs, and all other players have also played 8 games, and none have won more than 6 games, then the meta-game AI KNOWS that it is in the play-offs regardless of winning the next game. Therefore, it's 'win' state will be to develop it's players the best way possible in order to win the playoffs, and here, getting that 1 SPP on the Saurus could be very important in order to get it Block for the play-offs.

You could absolutely do that kind of thing, but this requires developing TWO distinct AI's playing two completely different games - one that is playing 'team manager' which was alluded to earlier in the thread, and one that actually plays the current game of bloodbowl.

The contempt factor is actually very similar to what I was suggesting with my difficulty modifier for the memory values in my first post. I would personally refine it a lot to also provide personalities, as opposed to purely difficulty, so that you COULD have an AI that liked to play passing dwarves, or a nurgle team manager AI that never takes claw or mighty blow.
Lorebass



Joined: Jun 25, 2010

Post   Posted: Jul 20, 2019 - 00:50
FUMBBL Staff
Reply with quote Back to top

Psh, set up an AI to watch semi-finals and better majors and minor tournaments. AI LEARNS
Roto



Joined: Oct 12, 2018

Post   Posted: Jul 20, 2019 - 01:03 Reply with quote Back to top

thoralf wrote:
Team management is indeed another task. Two AI could be needed. One would be the BB coach, the other would be the BB owner.

Wouldn't it be far simpler to just have a bunch of pre made teams at varying TV values for the bot to pull from instead of trying to build a bot to manage a team itself?

_________________
Image
thoralf



Joined: Mar 06, 2008

Post   Posted: Jul 20, 2019 - 01:59 Reply with quote Back to top

Roto wrote:
Wouldn't it be far simpler to just have a bunch of pre made teams at varying TV values for the bot to pull from instead of trying to build a bot to manage a team itself?

Of course! We could also add kick-off formations and all kinds of knowledge we don't need the engine to perfect.

Nelphine wrote:
Why would you need a 'witness' for the scores?

I thought your "current score for game size" referred to an accounting of the number of games. When one of the teams leads at 2-1 at T16 of one game, there's no need to count the OT for that game. Same for games that ends with cleared pitches. Once one tries to count deaths, stuns, knocked downs, KOs, Riots, Pitch invasions, or Swealtering Heat events, you need to memorize them.

Chess and Go are mostly eventless and memoryless, except for kos, castling, en passant, stuff that is easy to encode in the game representation. If we add historical details to how you identify what constitutes a game, BB gets massively huge. Just think of dice results. You have the very same game G1 and G2 up to T16, but then instead of scoring and winning on the 2+, you roll a 1 and draw. These become two different games!

Just imagine the odds that two games be similar up to T16!

***

The amount of such possibilities should not matter much from the bot's perspective. Under the machine learning paradigm, the bot learns by playing. It picks the characteristics that makes it better.

Here's how it could go. During its first thousand games, it may learn that blocking can be rewarding and blitzing each turn is better than not blitzing at all. Then it could learn not to start its turns with 1Ds for no good reason. As soon as it would discover that 2D rocks, its performance should improve. Then it could start to favor positions against which 2Ds are harder to get. Then it could find ways to pressure positions with Guard locks, to puts players in scoring position in case of fumbles, etc.

The magnitude of possibilities does not matter much when the AI learns to dismiss moves it finds irrelevant. Once it finds a good move, it plays it. If it's not a good move, it should adapt. Since it's dice-based, it may take time. But theorically it will happen nevertheless. At no point it needs to have a full list of all possible moves it could play or to take into account every single aspect of the situation.

Lastly, the more complex is your evaluation function, the more it costs in computation. This in turn may lower the quality of the evaluation. So there's an important trade-off. Bear in mind that the agent should in principle play its turns in 4 minutes, or else I know a few guys who will relish in timing it out!

_________________
There is always Sneaky Git.
Traul



Joined: Jun 09, 2013

Post   Posted: Jul 20, 2019 - 02:13 Reply with quote Back to top

Roto wrote:
Wouldn't it be far simpler to just have a bunch of pre made teams at varying TV values for the bot to pull from instead of trying to build a bot to manage a team itself?
If only there was a database of thousands of teams of all races at all TVs available somewhere...
Display posts from previous:     
 Jump to:   
All times are GMT + 1 Hour
Post new topic   Reply to topic
View previous topic Log in to check your private messages View next topic