5D Library › Bayesian Search in Unknown Landscapes, Optimal Stopping and Gittins Indices · 中文

Chapter 7 Reading the Classics: Victory Not Duration, Roads Not Taken, and Trial in Office

This chapter reads the research of earlier chapters beside several classical Chinese texts: the “Waging War”, “Nine Variations” and “Manoeuvring” chapters of the Sunzi, the phrase “speed is what the army prizes” (兵贵神速) from the biography of Guo Jia in the “Book of Wei” of the Records of the Three Kingdoms, and the “Eminent Teachings” and “Two Handles” chapters of the Han Feizi. The aim is not to show that the ancients already understood the theory of stopping times or multi-armed bandits, but to see how classical texts expressed the ideas that “delay has a cost”, “not every road must be taken”, “haste has a cost” and “judge by trial rather than appearance”, and to state three boundaries. First, these texts concern war and the arts of rule; their aims are victory and control, not maximising expected reward under uncertainty. Second, they are maxims drawn from experience, with no probabilities, priors or stopping rules, and cannot be matched one-to-one with any theorem. Third, the most important uncertainty in ancient warfare came from the enemy’s actions — a game rather than a contest against nature — which differs from the setting of the secretary problem and the bandit problem. All translations in this chapter are the author’s own.

Victory is prized, not duration

Source text

故兵闻拙速,未睹巧之久也;夫兵久而国利者,未之有也。……故兵贵胜,不贵久;故知兵之将,民之司命,国家安危之主也。(杨丙安 校理 1999)

(Sunzi, “Waging War”)

In the author’s translation: In war, one has heard of clumsy speed, but never seen skill that prolongs; there has never been a protracted war from which the state benefited. … Therefore in war victory is prized, not duration; the general who understands war is the arbiter of the people’s fate and the master of the state’s safety or peril.

Original meaning and context. “Waging War” is about the economic burden of war: earlier in the chapter, raising an army of a hundred thousand and carrying grain a thousand li costs “a thousand pieces of gold a day”, and a long war wears out the troops, drains the treasury and invites other lords to rise up. Sunzi’s “speed” thus comes from attrition: every extra day carries a certain and accumulating cost. “Clumsy speed” does not praise clumsiness; it says that when speed and finesse cannot both be had, choose speed.

Read beside modern research. In modern terms, “Waging War” stresses the cost of waiting. The discount factor gg of Chapter 4 and the cost per trial cc of Chapter 5 are both factors that make “one more observation” less worthwhile: the smaller gg and the larger cc, the earlier the optimal rule stops and the smaller the exploration bonus in the Gittins index. The interview-cost example of Chapter 6 shows the same structure: at a cost of 0.02 the optimal rule interviews only 5 applicants on average. But Sunzi does not discuss the benefit of waiting, the value of information; in the theory of stopping, the optimal stopping point is exactly where the marginal cost of waiting equals the marginal value of information. Taking only the “prize speed” half gives the kind of rule that “quits too early” in Chapter 5.

Speed is what the army prizes

Source text

至易,嘉言曰:「兵贵神速。今千里袭人,辎重多,难以趣利,且彼闻之,必为备;不如留辎重,轻兵兼道以出,掩其不意。」(陈寿 撰,裴松之 注 1982)

(Records of the Three Kingdoms, “Book of Wei”, biography of Guo Jia, juan 14)

In the author’s translation: When [the army] reached Yi, Guo Jia advised: “Speed is what the army prizes. Now we are striking at men a thousand li away; with so much baggage it is hard to seize the advantage, and when they hear of us they will surely prepare. Better to leave the baggage behind and send out light troops at double pace, to fall on them unawares.”

Original meaning and context. “Speed is what the army prizes” is often taken to come from the Sunzi, but it comes from a speech of Guo Jia recorded in the Records of the Three Kingdoms, during Cao Cao’s northern campaign against Yuan Shang and the Wuhuan of the three commanderies. Guo Jia gives two reasons: baggage slows the march and makes it hard to seize the moment; and, more important, if the army is slow the enemy will hear of it and prepare. Cao Cao followed the advice, led light troops out through the Lulong pass and defeated the Wuhuan.

Read beside modern research. Guo Jia’s reason for speed differs from Sunzi’s: Sunzi speaks of attrition, Guo Jia of an opponent who changes because you delay. In the secretary problem, Bayesian search and the bandit problem, “nature” does not change because you observe it; where there is an opponent, observation itself leaks information, and the opponent’s response lowers the value of waiting. This is exactly the nature of the game of googol that Ferguson pointed out: the person writing the numbers is an opponent (Chapter 1). When “speed is what the army prizes” is used as a maxim for stopping problems, one must therefore be clear whether the cost of delay comes from attrition, from the opponent’s response, or from opportunities slipping away.

There are roads one does not follow

Source text

途有所不由,军有所不击,城有所不攻,地有所不争,君命有所不受。……是故智者之虑,必杂于利害,杂于利而务可信也,杂于害而患可解也。(杨丙安 校理 1999)

(Sunzi, “Nine Variations”; the framework reads 涂 for 途, the two characters being interchangeable; the book follows the received text in reading 途)

In the author’s translation: There are roads one does not follow, armies one does not strike, cities one does not attack, ground one does not contest, and commands of the ruler one does not obey. … Therefore the wise in their deliberations always weigh benefit and harm together: weighing harm amid benefit, their undertakings can be relied on; weighing benefit amid harm, their troubles can be resolved.

Original meaning and context. “Nine Variations” is about how a general adapts to particular circumstances. The five clauses beginning “there are roads one does not follow” say that not every passable road, attackable enemy or assailable city is worth the effort; one must choose by benefit and harm. “Commands of the ruler one does not obey” says that a general in the field must decide by the actual situation rather than carry out orders mechanically. “Weighing benefit and harm together” means looking at both sides in every judgement.

Read beside modern research. These lines connect with the allocation rule of search theory: the greedy rule of Chapter 3 searches only a few cells at any moment, and cells whose product of probability and detection probability is too low are, for the time being, “not followed”. They also connect with successive halving in Chapter 6, where some lanes are stopped and tested no further. “Weighing benefit and harm together” connects with the one-armed bandit of Chapter 5: the value of continuing contains both the immediate expected reward (which may be negative) and the future value of information (which is positive), and looking at only one side leads to error. But the boundary is equally clear: Sunzi’s “not following” and “not striking” come from the general’s judgement of experience and give no calculable criterion; in search theory a cell “not searched for now” becomes worth searching again after others have been searched, and this dynamic reallocation has no counterpart in “Nine Variations”.

The indirect made direct, and marching with armour rolled up

Source text

军争之难者,以迂为直,以患为利。故迂其途,而诱之以利,后人发,先人至,此知迂直之计者也。……是故卷甲而趋,日夜不处,倍道兼行,百里而争利,则擒三将军,劲者先,疲者后,其法十一而至。(杨丙安 校理 1999)

(Sunzi, “Manoeuvring”)

In the author’s translation: The difficulty in contending for advantage lies in making the indirect route direct and turning misfortune into gain. So take a roundabout road and lure the enemy with a bait; set out after him and arrive before him: this is to know the stratagem of the indirect and the direct. … Therefore, if one rolls up the armour and hastens, halting neither day nor night, marching at double pace, and contends for advantage a hundred li away, the commanders of all three armies will be captured: the strong will arrive first and the weary fall behind, and as a rule only one in ten will arrive.

Original meaning and context. “Manoeuvring” concerns the contest for the initiative. It recommends gaining the initiative by indirect routes, and warns of the cost of forced marches: haste scatters and exhausts the army, so that only a small part actually arrives.

Read beside modern research. The passage is a reminder that “fast” has a cost too: setting the boundaries of the sequential probability ratio test of Chapter 5 close together for speed means accepting higher error rates. The trade-off between speed and accuracy can be calculated in the theory of stopping; in “Manoeuvring” it is a warning from experience, and “one in ten” is hyperbole, not a statistic.

Try them in office and assess their accomplishments

Source text

夫视锻锡而察青黄,区冶不能以必剑;水击鹄雁,陆断驹马,则臧获不疑钝利。发齿吻形容,伯乐不能以必马;授车就驾,而观其末涂,则臧获不疑驽良。观容服,听辞言,仲尼不能以必士;试之官职,课其功伐,则庸人不疑于愚智。故明主之吏,宰相必起于州部,猛将必发于卒伍。(陈奇猷 校注 2000)

(Han Feizi, “Eminent Teachings”)

In the author’s translation: If one looks only at the tin added in casting and at the bluish or yellowish hue of the metal, even Ou Ye could not be sure of a sword; but strike at swans and geese in the water and cut down colts on land, and even a slave will not mistake blunt for sharp. If one opens the horse’s mouth to look at its teeth and examines its shape, even Bo Le could not be sure of a horse; but put it to a carriage and watch it to the end of the road, and even a slave will not mistake a nag for a good horse. If one looks at a man’s face and dress and listens to his words, even Confucius could not be sure of a scholar; but try him in office and assess his accomplishments, and even an ordinary person will not mistake the foolish for the wise. Therefore among the officials of an enlightened ruler, the chief minister must rise from the provincial posts, and the fierce general must come up from the ranks.

Original meaning and context. “Eminent Teachings” criticises the Confucians and Mohists for “judging people by appearance” and “judging people by words”, and argues for identifying talent by actual trial and assessment. The three comparisons share a structure: judging from appearance, even the best expert cannot be sure; judging from the results of actual use, an ordinary person can see clearly. The last sentence draws the institutional conclusion: senior officials should be selected step by step from the lower levels.

Read beside modern research. The passage corresponds clearly to several themes of this book. “Looking at face and dress and listening to words” is like a prior based on weak signals, and “trying him in office and assessing his accomplishments” is like updating one’s judgement with real trial results: Chapter 3 says the same thing, that the quality of a judgement depends on the “detection probability” of the evidence, and appearance and words detect ability poorly. “The chief minister must rise from the provincial posts” is a form of stepwise screening: try people first in low-risk positions, and let only those who do well move up, a structure like successive halving in Chapter 6.

But the boundary is also clear. Han Fei’s trial is not a randomised experiment, and it considers neither the cost and risk of the trial itself nor whether performance “in the provinces” predicts performance “as chief minister” — exactly the questions that modern research on personnel selection tries to answer.

Examining whether forms and names tally

Source text

人主将欲禁奸,则审合刑名者,言异事也。为人臣者陈而言,君以其言授之事,专以其事责其功。功当其事,事当其言,则赏;功不当其事,事不当其言,则罚。(陈奇猷 校注 2000)

(Han Feizi, “Two Handles”; the received text reads 言异事也, which some critical editions emend to 言与事也; the translation follows the emended reading)

In the author’s translation: If the ruler wishes to prevent wrongdoing, he examines whether “forms” (deeds) and “names” (words) tally — that is, words and deeds. A minister sets out his proposals; the ruler assigns him tasks according to his words and holds him to account solely by the results of those tasks. If the results match the task and the task matches the words, he is rewarded; if the results do not match the task, or the task does not match the words, he is punished.

Original meaning and context. “Two Handles” is about the ruler’s two handles of control over ministers, punishment and reward. “Examining whether forms and names tally” requires ministers to state in advance what they will do and can achieve, and to be assessed afterwards against that commitment. Chapter 7 of Volume 27, Mechanism Design, Incentive Compatibility and the Game Theory of Negotiation, has already discussed how this passage punishes both overstatement and understatement; it is recalled only briefly here.

Read beside modern research. From the standpoint of this book, the point of “examining whether forms and names tally” is the standard written down in advance: assessment is based on what was said beforehand, not on explanations afterwards. In form this resembles the stop-loss boundaries set in advance in Chapter 5 and the pre-registered trial of Chapter 8: write down the rule of judgement before seeing the results. But Han Fei’s purpose is the ruler’s control over ministers, and his “punishment” is penal, not the abandonment of a project; modern pre-registration binds the researchers themselves and is not a means of punishing others.

Three boundaries

First, all these texts were written in the context of war and rulership, with victory and control as their aims, not the maximisation of expected reward under an explicit objective; to read them as “ancient decision theory” is anachronistic. Second, they have no concepts of probability, prior, threshold or stopping rule, and no calculation of any kind; the “correspondences” are only resemblances of structure. Third, the most important uncertainty in ancient warfare came from the enemy, and Guo Jia’s “speed” was aimed precisely at the enemy’s response; whereas most conclusions of the secretary problem, search theory and the bandit problem assume that “nature” does not respond to one’s actions. Problems with an opponent need game theory, not the theory of stopping for a single decision maker.

Summary

“In war victory is prized, not duration” in “Waging War” stresses the cost of waiting; “speed is what the army prizes” in the biography of Guo Jia stresses that the opponent changes if one delays — two different reasons. The theory of stopping also counts the benefit of waiting, the value of information. “There are roads one does not follow” and “weighing benefit and harm together” in “Nine Variations” connect with the allocation rule of search and the trade-off between continuing and abandoning, and “Manoeuvring” warns that haste too has a cost. “Try them in office and assess their accomplishments” in “Eminent Teachings” argues for judging talent by the results of trial rather than appearance, and “the chief minister must rise from the provincial posts” is a form of stepwise screening; “examining whether forms and names tally” in “Two Handles” stresses a standard written down in advance. The ancients did not “already discover” the theory of stopping; the resemblance between these texts and modern research lies in the structure of the problems, not in the methods. The next chapter gives the checklist and the trial protocol.

References

杨丙安 校理. 1999. 十一家注孙子校理. 中华书局.
陈奇猷 校注. 2000. 韩非子新校注. 上海古籍出版社.
陈寿 撰,裴松之 注. 1982. 三国志. 2nd ed. 中华书局.

Fan, Y. (2026). Bayesian Search in Unknown Landscapes, Optimal Stopping and Gittins Indices: Exploration–Exploitation Trade-offs and Sunk-Cost Truncation in Statistical Decision Theory [未知地形中的贝叶斯搜索、秘书问题与吉廷斯指数] (Version 1.0.0). 5D Library. https://doi.org/10.5281/zenodo.23810102 · Licensed under Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0).