Григорий (gest) wrote,

Цитаты и пометки на полях

Да, меня тут спросили, откуда я черпал информацию для предыдущего поста. Всё очень просто, в сети лежит текст самого Лената, "EURISKO: A program that learns new heuristics and domain concepts". Чтобы вы не думали, что я это всё высосал из пальца, вот куски, на которые я опирался:

"...Why was the Worth of EnergyGun lowered? Through many tens of simulations, it became clear that one could buy enough armor plating to make a ship invulnerable to attacks by these types of weapons, and from then on almost all ships were so armored.

Thus, any ships having energy weapons were at a serious disadvantage, and gradually—as they lost—the Worth of EnergyGun declined. Incidentally, this mistakenly led to a correct heuristic: "If a weapon cannot hit at all at one range (e.g., Long Range in this case), then it's probably not worth having too many of them." That was not the major problem with energy guns, but the heuristic is a good one anyway."

"If more than one proposed heuristic remained, new variant fleets were designed and simulated, each one embodying one but violating the other heuristics. In cases of circular victories (A beats B who beats C who beats A) all the candidates involved were retained, but with somewhat lowered worth. Such situations were interpreted as analogues of local maxima, and EURISKO would try for a very different fleet design for the next iteration."

"So fleets fight (each battle taking between 2 and 30 minutes), and the battle is analyzed to determine which design policies are winning, and occasionally what fortuitous circumstances can be abstracted into new design heuristics. An example of the former (gradual parameter adjustment) was when the Agility of ships gradually decreased, in favor of heavier and heavier Armor plating of the hulls. An example of the latter (fortuitous monsters) was when a purely defensive ship was included in an otherwise-awful fleet, and that fleet could never be fully defeated because that defensive ship, being very small, unarmored, and super agile, could not be hit by any of the weapons of the larger nearly-victorious fleet."

"EURISKO also had many `mutation' operators, such as changing the number of ships, their size, their weaponry, etc. The many constraints—the TCS rules and formulae—were used to constrain the generation of mutant fleets, and to prune away illegal ones before simulating them. At first, mutations were random. Soon, patterns were perceived: more ships were better; more armor was better; smaller ships were better; etc. Gradually, as each fleet beat the previous one (and a few random ancestors), its "lessons" were abstracted into new, very specific heuristics.

One very general result that EURISKO abstracted from this evolutionary design process was a `nearly extreme' heuristic.

In almost all Traveller TCS fleet design situations,
the right decision is go for a nearly—but not quite—extreme solution.

Thus, the final ships had Agility 2 (slightly above the absolute minimum), one weapon of each type of small weapons (rather than 0 or many), the fleet had almost as many ships as it could legally have but not quite (96 instead of 100), etc.

Big weapons (enormous spinal mounts capable of blasting another ship to pieces with a single shot) were gradually phased out, in favor of an enormous number of small missile weapons. The fleet had almost all (75) ships of this type though there was one ship which was small and super agile and purely defensive (and literally unhittable by any reasonable enemy ship), and a couple monstrous hulks which had no chance of defense against normal ships, butwhich had weapons just barely accurate enough to hit any enemy ships that were (of course!) small and agile and purely defensive.

Some of the strangest elements of the final fleet were discovered accidentally rather than as the result of a long, continuous evolution process. The usefulness of a tiny defensive ship was apprehended after a `lifeboat' was the only survivor from one side's fleet, yet round after round it could not be hit at all. That design was immortalized into a design strategy ("Include one such ship in your fleet!"), and a very general rule began looking for ships that could destroy it. Finally, one was found; it was quite strange, and would never have been included except to counter the possibility that the enemy might have small defensive ships too. Against any normally-armed ship, it would quickly be destroyed. Basically, this new ship had moderate size, no armor, the largest possible guidance computer, the slowest possible engines for its size and equipment, and one single, enormous accelerator weapon — a weapon usuall yignored because its broad beam glances harmlessly off large armor-plated ships, but which is very easy to aim. This combination is ineffective for most combat, but is just enough to fire at the little boats it might be sent against. We were a little disappointed that none of the other entrants had small defensive "stalemate guarantors" of the sort we took.

Almost all the other entrants in the final tournament had fleets that consisted of about 20 ships, each with a huge spinal mount weapon, low armor, fairly high agility, and a large number of secondary energy weapons (laser-type weapons). This contrasted with EURISKO'S fleet in almost all ways.

Most ships in our fleet did sprout one solitary laser among their 50 or so weapon batteries, but not because it was useful in combat — just to absorb damage from enemy fire (thanks to the somewhat unrealistic scheme by which damage is inflicted on ships which have been hit). After an exchange of fire, most of the enemy behemoths did indeed sink one of EURISKO'S ships, for a total loss of about 15 ships. In return, EURISKO'S 96 ships sank about 5 of the enemy. So just prior to the second exchange of fire, the enemy was down to 15 ships, and EURISKO 81. After a second round of fire, the numbers were 11 and 70. Two more exchanges brought the totals to 1 and 46, and one more round of fire wiped out the enemy. In this scenario—which was the most common one in all EURISKO'S battles during the tournament — there is no need at all to bring any of it's specialty ships into the front lines at any time.

EURISKO's few specialty ships  remained unused until the final round of the tournament, battling for 1st versus 2nd place . That opponent also had ships with heavy armor, few large weapons, low agility, etc. He was lacking any fast ships or fast-ship-killers, though. The author simply pointed out to him that if EURISKO were losing then (according to the TCS rules) our side need put only our fast ship out the front line, withdraw all the others and repair them, and—once they were finished repairing themselves — effectively start the battle all over again. This could go on ad infinitum, until such time as EURISKO appeared to be winning, and in that case we would let the battle continue to termination. The opponent did a few calculations and surrendered without fighting. Thus, while most of the tournament battles took 2–4 hours, most of those involving EURISKO took only a few minutes.

The tournament directors were chagrined that a bizarre fleet such as this one captured the day, and a similar fleet (though not so extreme) took second place. The rules for future years' TCS tournaments were changed to eliminate the design singularities which EURISKO found. For example, repairing of damaged ships was prohibited, so the utility of the unhittable ship became negligible.

When rules for the 1982 tournament were announced, EURISKO was set to work on finding a new fleet design. Although many of its best designs and design rules were now illegal or useless, most of the general heuristics it synthesized about the game were still valid. Using the `nearly-extreme' heuristic, for instance, it quickly designed a ship with practically no defense, and that ship filled a key role in the final fleet. Coincidentally, just as the defensive ship made a difference in the 1981 final round, the offensive ships made a difference in the 1982 final round. In each case, their presence caused the opponent to resign without firing a shot. The bulwark of our 1981 fleet was a ship that was slow and heavily armored; the majority of ships in our 1982 fleet were very fast and completely unarmored. Just as most `experienced' players jeered at the 1981 fleet because it had practically no large weapons, they jeered at the 1982 fleet because it was unarmored and it still had no large weapons, even though the rules changes had made them much cheaper."

"What EURJSKO found were not fundamental rules for fleet and ship design; rather, it uncovered anomalies, fortuitous interactions among rules, unrealistic loopholes that hadn't been forseen by the designers of the TCS simulation system. There may be little of what EURI5KO found that has application to real naval design ; most of its findings pertained to the fine structure of the TCS rules, not to the real world. For example, a crew hit reduces the number of crewmen on a ship from n down to the largest power of 10 smaller than n (e .g., from 370 to 100, from 82 to 10); EURISKO therefore designed ships requiring 99 crewmen, and crewed them with 101 people; the first crew hit therefore had no effect on the ship's battleworthiness.

The fact that EURISKO's discoveries were synergistic loopholes rather than genuine naval insights is not in itself bad, as our goal was to win the tournament, not break new ground in real warfare . In fact, the very unreality of the TCS rules — as any 100-page model of the real world is bound to be incomplete and have rough edges—promised to aid us in our task. Here was a search space that had not been explored much by human beings yet; most designers were applying analogues of rules that hold in real life, and that yielded them reasonable designs — fleets of the kind the TCS people anticipated. EURISKO was able to walk around in the space defined by the set of rules, somewhat awkwardly, but (thanks to its absence of common sense knowledge) with few preconceptions about what an optimal design might be. Perhaps we will know that the program has `arrived' when it first fails to win the TCS tournament. This notion of a large, unexplored search space, not necessarily well-matched to our everyday comnon-sense intuitions, will come up again and again in the following pages. It appears to characterize those domains for which automated discovery (of both concepts and heuristics) is currently most viable."

И раз уж об этом зашла речь...

Во-первых, по словам организаторов турнира, они не забанили Лената и не принудили его к самомодерации, а присвоили ему почётный пожизненный титул "Гранд-Адмирала Флота" и вежливо попросили уйти непобеждённым и дать поиграть малькам.

Во-вторых, это не очень понятно из самого текста, но на самом деле, в финальном флоте Лената образца 81 года было не три типа кораблей ("черепашка", "шлюпка", "снайпер"), а шесть. "Черепашек" с ракетами там было всего 75 из 98 - три четверти. Класс "неубиваемый нормальными средствами корабль" состоял из двух типов кораблей - корабля минимального размера и более крупного (но всё ещё очень маленького) корабля с максимальной скоростью, и в составе эскадры их таких было несколько штук. Три корабля были снайперами - охотниками на быстрые мелкоразмерные цели. Для чего были нужны два других типа корабля, наверное, уже и сам Ленат не помнит.


Приколы, которые упоминает Ленат, но которые не вошли в мой текст.

На самом деле, фишка со "шлюпкой" состояла в том, что по изначальным правилам турнира, часть кораблей можно было уводить в резерв и там чинить. Таким образом, Ленат мог взять измором любого противника, сделавшего ставку на обмен ракетными ударами - минимальные неубиваемые корабли могли заблокировать целый вражеский флот, пока остальные корабли отходили в тыл и восстанавливали свои повреждения.

Отдельная ржака была связана с принципом "критических попаданий по жилым отсекам". Правила были рассчитаны на экипаж в несколько сотен (или даже тысяч) человек, и гласили, что в случае критических попаданий количество людей на борту откатывается к следующей степени десятки - с 2000 до 1000, с 300 до 100 и так далее. Что, естественно, вынуждало игрока распределять выживших по боевым станциям и оставлять какие-то части корабля без персонала. Eurisco сразу забила на эту игру в сериал "Звёздный путь" - её небольшим кораблям ("черепашкам") требовалось около ста человек для обслуживания всех ракетных батарей, поэтому они строились в расчёте на 101 члена экипажа, где сто первый был лишним. Первый "крит по экипажу" не влиял вообще ни на что: команда из сто одного человека превращалась в команду из ста человек. "Отряд не заметил потери бойца", воистину.

...Как я уже сказал, мне в этом тексте было нужно ровно одно правило, вот это:

"If more than one proposed heuristic remained, new variant fleets were designed and simulated, each one embodying one but violating the other heuristics. In cases of circular victories (A beats B who beats C who beats A) all the candidates involved were retained, but with somewhat lowered worth. Such situations were interpreted as analogues of local maxima, and EURISKO would try for a very different fleet design for the next iteration."

Но другое правило там тоже интересное:

"In almost all Traveller TCS fleet design situations, the right decision is go for a nearly—but not quite—extreme solution".

Просто была одна легендарная дискуссия с makarovslava, когда он сказал, что если уж делать в воздушном бою ставку на стэлс, то это должна быть абсолютная ставка на стэлс, никак половинчатых решений, никакого "а ещё он сможет летать, как почти нормальный самолёт", "а ещё тут можно снаружи бомбы и ракеты подвешивать". Если мы выбираем радикальное решение, мы должны довести его до конца. Отсюда следующее высказывание: В общем, перефразируя одну старую реплику makarovslava, проблема реального советского флота была в том, что он был недостаточно радикальным в своей "советскости". Под давлением внешних обстоятельств была выбрана стратегия асимметричного ответа, но последовательность в её проведении отсутствовала.

Соответственно, в мозгах Eurisco этому соответствовало правило, что оптимальным решением скорее всего окажется почти предельное решение, почти - но не совсем. Если мы экономим на скорости, то она должна быть почти минимальной, если мы играем от численного превосходства, количество кораблей должно быть почти максимальным, и так далее.


Забавная деталь - о турнире 82 года практически ничего неизвестно, кроме факта самой победы Лената, и того, что организаторы слегка видоизменили правила, заставив игроков сделать упор на скорость. В итоге, тактика Eurisco строилась на подрыве собственных повреждённых кораблей, чтобы всё время удерживать среднюю скорость эскадры в районе максимально возможных показателей.


Мой пост восприняли, как наезд в адрес Розова. Скажем так, суть наезда в него так и не вошла.

Но не удержусь :). Да, дико смешно, что Розов не понимает разницу между войной на море и космооперой, реальными условиями и условными правилами. Но главный его косяк в следующем:

"Пикантность ситуации состояла еще и в том, что флотилия в стиле EURISCO (будь она построена реально, а не виртуально) стоила бы в сотни раз дешевле, чем традиционно-организованные флотилии, которые она побеждала в штабных играх".

Пикантность ситуации состояла в том, что всем участникам выдавался один и тот же бюджет, в этом состоял главный принцип состязания. То есть, если считать, что результат Лената имел хотя бы малейшее отношения к принципам войны на море, то он означал следующее - если эскадра боевых кораблей заплывёт на минное поле (а малоподвижные ракетные "черепашки" Лената вполне тянули на противокорабельные мины настоящего и будущего), и при этом само минное поле будет стоить, как полноценная эскадра, то этой эскадре конец.

Остальное - это уже мелочи. Но да, я бы хотел прочитать у Розова о том, как вдохновлённые результатами Eurisco меганезийцы строят свою стратегию вокруг скоростной безоружной моторки, по которой "нельзя попасть", рассчитывая, что она надёжно затормозит продвижение вражеского флота (они же не смогут в неё попасть! она маленькая и быстрая!). И берут на борт своих боевых кораблей хотя бы одно Ненужное Оружие, чтобы, когда враги начнут по ним стрелять, они попадали в Ненужное Оружие вместо нужного. И, конечно же, меганезийцы должны обязательно включать в состав команды одного Лишнего Человека (а что? будут стрелять, его первым убьют, выгода!).
Tags: космоопера
  • Post a new comment


    Anonymous comments are disabled in this journal

    default userpic

    Your reply will be screened

    Your IP address will be recorded