by Peter Byrne
Rather than try to keep track of every possible weapon-target permutation, an incomputable task, QUICK randomly sampled the vast range of probable outcomes to select the most probable results. And without this automatic sampling method, there would have been no solutions, as trying to solve these problems one step at a time—linearly–was technologically impractical. It took tremendous skill to operate the QUICK simulation. Every time that the computer made another pass over the data the program learned from the experience, gradually optimizing the deployment of weaponry. Part of the high art involved from the human end was originally assigning values to the multipliers in such a way as to avoid becoming stuck in recursive feedback loops.8
Lambda’s Paul Flanagan was the project director for QUICK in 1969. He programmed it through several generations of missile systems, including Polaris and MIRVed Minuteman missiles (MIRV stands for “multiple independently targetable reentry vehicles”). He says that QUICK mirrored the SIOP, so it had to access the same databank of information about targets, damage functions, kill ratios, etc. It was constantly updated with hard intelligence on the capabilities and locations of American and Russian nuclear forces. But it was not the SIOP, i.e. Lambda staff did not know the precise content of the targeting plan at any given moment.
QUICK also ran its twin: the RISOP program, which was the mirror of the Soviet version of a SIOP. Flanagan says that Everett once came up with a RISOP configuration that completely wiped out America. “If the Russians get this, we are toast,” Everett told his colleagues.9
Kill probabilities
The Advanced Research Projects Agency wanted to know how many warheads should be dropped on an urban area when the exact location of a target, such as a steel factory, or a radar installation, a government building, or a food supply warehouse, was uncertain. Demonstrating how to do this, Everett and Galiano authored, “Some Mathematical Relations for Probability of Kill.” Using the multipliers, the idea was to figure out the probability that a random point can be destroyed by aiming a variable number of weapons at the city center, with civilians to be treated as “collateral damage”.
After parameterizing constraints such as aiming errors, dud possibilities, air or ground burst choices,
The remaining step in the calculation of expected damage to the city is then a summation or integration of the damage at all points in the city weighting each point by its value. Thus, the focus enlarges from individual buildings to all the buildings in the city, finally encompassing all structures or people in the city…. We shall accordingly adopt this general function as our destruction function, subject to subsequent empirical verification.10
Galiano wrote a related paper on how to defend cities against ballistic missile attacks that include decoys. He addressed the overriding problem with ABM planning (then and now): it is difficult to justify deploying an expensive ABM defense system because it costs far less to fool it with decoys or to overcome it with a shower of missiles than it costs to build and maintain it. The upshot of Galiano’s paper was that by using a “preferential defense,” where selected targets within a region are preferentially defended, and the enemy does not know which ones will be defended, an ABM defense could be cost-effective.11
The next year, Everett wrote a game theoretic paper showing the ABM concept to be ineffectual when there is a limited supply of missile interceptors—the enemy can just keep shooting at the target until the defender’s supply is exhausted. The problem in defense is, of course, that all it takes is for one missile in a salvo to penetrate the defense and the ABM radar and interceptors are history. Everett concluded that “the defense is in a quandary,” because if it shoots at all incoming missiles it will exhaust interceptor supplies and fall prey to the next wave of attack. But if it holds back, the chances of an incoming missile penetrating the defense vastly increase.12 This conclusion is common sense, but the Pentagon often paid experts to back up common sense with expensive calculations, and billions of dollars were at stake.
Modeling possible deployments of blast and fallout shelters as a cost function of lives to be saved was another bread and butter job. For the cheapest shelter, the cost of a life came in at $194, rising to $381 for providing a cozier, harder habitat; the premium shelter cost $38 billion for 100 million lives (five percent of the gross domestic product). Lambda then compared the strategic benefits of protecting industry from blast damage versus not building shelters for the population. In the end, massive sheltering was shown to be worthless if it resulted in the enemy deploying bunker-busting hydrogen bombs.13 Shelters were never built in any meaningful way in the United States, although Everett kept his ad hoc basement shelter stocked with canned food and guns, just in case.
The game of sex
Flanagan remembers Everett, Dean, Pugh, and Galiano obsessively playing Kriegsspiel chess in a room set aside for that purpose. Kriegsspiel was game theory brought to life and nobody played regular chess at Lambda. In Kriegsspiel, each player has a separate chess board. Neither player can see his opponent’s board. A referee announces whether proposed moves are legal or illegal. After a while, the game had to be banned from the office, as it was cutting into billable man-hours.
“Everett was the smartest man I ever met,” says Flanagan. “He was an incredible out-of-the-box thinker. But he had a terrible attitude towards women. He treated them as objects. He would have affairs in the office, and there would be fallout and the women would quit.” But that did not bother the boss.
Hugh viewed life as a mini-max game. Unfortunately, his objective function didn’t include emotional values. I don’t think he meant to be bad to people, he just did not think about the emotional impact of what he did on others. He was not warm or friendly. He hurt many people by how he treated them. And he drank too much.
We talked about his multiple universe theory once. There seemed nothing wrong with it except it was a bit goofy.
Flanagan quit Lambda in 1971 because it was starting to bother him that “targets” were, in fact, people. He no longer wanted to spend his days thinking about whether it was more cost-effective to kill 160 million or 180 million people. There were too many fingers on the trigger, and an accidental holocaust could be caused like the firing of the first shot at Concord, Flanagan worried. So, looking for a less explosive solution to the world’s problems, he took a job at the Christian Broadcasting Network for, “the second smartest man I ever met: Pat Robertson.”14
Although Everett relished his role as the CEO of an important company, he left the day-to-day operations to Dean and Pugh and Killalea. But his sexism trickled down. Killalea often hired “Miss America-type” secretaries, and then hit on them for sexual favors.15
Everett’s co-workers do not remember him slurring his words, or being falling down drunk, but it was obvious to many of them that his heavy drinking was debilitating. Returning to the office after lunch, he would take a nap, arising refreshed and productive. But he gradually became pasty-faced and obese; his breath reeked of alcohol, his clothes of tobacco smoke; he became increasingly intolerant of people he considered to be unintelligent. And even as the feminist movement gathered steam in American society, he gleefully treated women as sexual objects, which repulsed some of his colleagues.
“He attacked women on a mental, not a physical level,” recalled one of his closest friends, Don Reisler. “Think about the Selfish Gene and the fact that men and women operate differently and somehow he convinced them that having sex was the logical thing to do. Plus, he spent a lot of money and seemed to know a great deal about the fine points of life.
“Plus, he was really quite charming.”16
Gender wars
Born in 1928, Joanna Frawley was a first-generation operations researcher and computer programmer. With her Q-clearance at WSEG and, then, Lambda, she worked with Everett on weapons systems for many years.17
The office atmosphere at the Pentagon in the early 1950s was treacherous for a woman: “I initially did research papers for a man who sim
ply erased my name and put his own name on them.” But Frawley’s brains could not be stolen, and by the time Everett arrived, she was appreciated for her skills. “Those were the days when computers and programmers were rare, so everything you did was extraordinary. Everything had to be written from scratch.” She quickly rose to be WSEG’s Chief Applications Programmer. She modeled aircraft designs, programmed nuclear war gaming systems, and built nuclear blast and fallout damage models. One of her first big projects was writing a program to graph the lethal relationship between megatonage, fallout patterns, and radiation dosages. Recognizing her knack, Everett included her in his Strategic Analysis Group, which spearheaded the statistical research for Report 50, and she became proficient at wielding his multipliers.
Frawley remembers her boss as brilliant, but sad. And he did like his cocktails. “Everybody smoked and went to poker and dinner parties in each other’s homes where we drank martinis and old fashioneds.” And there was wife-swapping. “This was the LSD era when people were becoming more emancipated and wilder and an awful lot of things went on.” But Frawley had her limits: “I would never have had an affair with Hugh because I found him to be a physically unattractive, smoking drunk. Not that he didn’t try, but he never pursued it after the ‘no.’ Obviously there were no hard feelings, because I worked for him many years.”
30 The Bayesian Machine
Science, business, politics, have lost all foundations and proportions which make sense humanly. We live in figures and abstractions; since nothing is concrete, nothing is real. Everything is possible, factually and morally. Science fiction is not different from science fact, nightmares and dreams from the events of next year. Man has been thrown out from any definite place whence he can overlook and manage his life and the life of society. He is driven faster and faster by the forces which originally were created by him. In this wild whirl he thinks, figures, busy with abstractions, more and more remote from concrete life.
Erich Fromm, 1955.1
Nuclear uncertainty
In early September, 1969, the United States struck the Soviet Union with 3,700 nuclear warheads exploded on 1,300 targets delivered by a combination of five types of aircraft and three types of missiles. Some of the weapons experienced unexpectedly high failure rates causing the “catastrophic failure of the war plan.”
Fortunately for the planet, the attack was a QUICK simulation aimed at teasing out hidden variables in destruction scenarios. From the computer game, Lambda determined that “overkill” might be wasteful in some circumstances, but, overall, it provided a “hedge against uncertainty.”2 Overkill capacity gave the military confidence in the war plan.
The QUICK study was based on “frequentist” probability theory. Frequentism is tied to how often an event has occurred in a series of experiments (i.e. relative frequencies of past events). But frequentism limits the scope of war games, because it is not possible to determine relative frequencies by experimenting with nuclear wars. As the sixties came to an end, Everett turned to another type of probability theory, called Bayesian inference, to glimpse the future.
Thomas Bayes was an 18th century mathematician who formulated an evidential theory of probability. The advent of “Bayes Rule” set off two centuries of debate on the fundamental nature of probability, which, despite its reliability as a calculational tool, remains ephemeral, abstract, almost indescribable.
Frequentists assert that probability has meaning only as a measure of the relative frequency with which a particular event has occurred over time: probability is frequency. Frequentists focus on events, such as the flip of a coin, that can be repeated many times. They assert that the probability of a particular outcome occurring—heads, say—is the number of times that heads has reportedly occurred in the past: one half of the flips.
Bayesians, by contrast, assert that probability is more than relative frequency. It is a function of a belief that an event will reoccur with a certain frequency. They measure probability by using Bayes’ rule, a formula that updates a prior belief about the probability of a recurring event with new information, new evidence in addition to relative frequency records. The new information might be the frequency with which similar events have occurred, or new data about changed conditions in the environment. For Bayesians, rational people make decisions based upon evidence-based beliefs about the likelihood of an event occurring. A frequentist might not find this method to be rational—but we will leave that debate to the philosophers.
J. P. Morgan gets mad
The construction of what Everett called his “Bayesian Machine” began after John Barry, a former member of WSEG, went to work for J.P. Morgan & Co. in New York. In 1971, Everett had a remarkable idea for building a computerized stock market timer; Barry convinced Morgan to front Lambda $10,000 to develop it. The prototype program, called Predictor, strove to optimize portfolio returns by quickly identifying new trends in financial markets according to Bayesian probability theory.
The program calculated the probable paths of stock prices as a function of current prices and historical trends. Like QUICK, Predictor reduced the computational scope of the problem at hand by selecting statistical samples, in this case from an enormous, branching universe of probable price paths. But Predictor added a new twist: it used Bayes rule to strip away the most improbable paths and uncover the presence of pricing trends, rising or falling. It aimed at unveiling turning points in market trends in time for investors to react.
There is, of course, no such thing as a foolproof stock market timer; the market is incredibly complex and random variables such as weather, psychology, and politics are driving factors in performance. Barry recalls that Everett’s timer was efficient, as far as timers go, but, “He refused to disclose the method to Morgan. He refused to give us the computer code and insisted that Lambda be paid for market forecasts.” Barry was very angry and considered suing Lambda after Everett started using the innovative program to do military research. He held his grudge for decades, saying, in 2002, that Everett was, “a brilliant, innovative, slippery, untrustworthy, probably alcoholic, individual.”3
Barry needn’t have worried; Everett was unable to make Predictor generate better than average returns on his own half million dollar portfolio.4 But in the form of a “Bayesian Machine” or “E-Filter,” it generated another kind of profit.
Filtering many worlds
Although the E-Filter flopped at outsmarting the Dow Jones average, it proved useful in operations research. In 1973, Lambda physicist Gary Lucas wrote “Cassandra, A Prototype Non-Linear Bayesian Filter for Reentry Vehicle Tracking.” Thirty-five years later, he enthusiastically diagramed his missile tracking system on a white board in his home office, which overlooks a lake in a forested area near Fairfax, Virginia. Talking to Lucas about operations research is the next best thing to talking to Everett himself, as the two men were intellectually in tune and very close.
Lucas spent part of his childhood at Los Alamos, where his father was a machinist working on the Manhattan Project. He came to Lambda in 1968 with a freshly pressed doctorate in theoretical physics from Yale. In one of their first conversations, Everett asked him to solve a series of tricky mathematical puzzles, and was delighted when he did so with alacrity. “Life was a game to Hugh,” says Lucas, echoing so many others.
Lucas says that Everett ignored men with whom he could not compete intellectually; and women were simply prey. For example, on a business trip to Puerto Rico, he propositioned an employee with whom Lucas was having an affair. Frantic to escape her amorous boss, she telephoned Lucas, who flew overnight to Puerto Rico, joining her in time for breakfast. When Everett walked into the restaurant for his omelet, he did a double take and grinned at the lovers—win some, lose some.
On Fridays, after work, Lambda traditionally hosted a sherry hour for employees and clients that often lasted past midnight. Lucas spent many an hour sipping sherry and chatting with Everett about operations research. Although they did not talk about his many w
orlds theory, Lucas says that he has a good idea of how his friend felt about it:
He thought that what he did was cool beyond belief, outrageously outside the nine dots. It tickled him to the core that it was at odds with what everyone had thought at the time. It challenged the authorities in physics, which gave him a quiet satisfaction. And it provided a simple, unarguable explanation for a fundamental problem in physics, which pleased him intellectually. He didn’t know if it was true, and to a considerable degree he didn’t care. It was all great fun. But when Bohr and friends strongly opposed the theory, he became invested in it and identified with it. He was not used to losing. He was also not used to living in the real world. It had all been a game for him. He couldn’t handle that. He broke it off and started a new life.
Cassandra calling
During the Cold War, scientists were (and remain) daunted by the complexity of shooting at incoming ballistic missiles. Using radar and computers to accurately predict the continuous trajectory of a missile speeding thousands of miles an hour with enough accuracy to destroy it with another speeding missile seemed impossible. First, the incoming missiles did not fly linearly, i.e. they could be programmed to weave and duck and dodge and launch multiple warheads and scatter decoys. Second, in Everett’s day computers did not possess enough memory and brain power to calculate all of the alternative trajectories—deviations from a smooth flight path—in time to launch an interceptor. Prior to Cassandra, predictive tracking filters were linear-thinking machines, hence they were practically useless when faced with maneuvering aircraft or flocks of reentry warheads. Assembling Cassandra from the corpse of Predictor, Lucas connected the problem of tracking average fluctuations in stock prices to the problem of predicting missile trajectories.
