Darwin Among the Machines

by George B. Dyson

  Although operating on principles similar to those of the Institute for Advanced Study, RAND went about the business of creative thinking in reverse. The Institute, established for the pursuit of pure research, quietly facilitated military work. RAND was openly targeted at military objectives, while quietly facilitating pure science and advancing scientific careers. Cross-fertilization between pure and applied mathematics flourished at RAND as nowhere else. In 1955, RAND published A Million Random Digits with 100,000 Normal Deviates, whose introduction notes that “because of the very nature of the tables, it did not seem necessary to proofread every page of the final manuscript in order to catch random errors.”40 In the 1950s there was a serious shortage of random numbers, and RAND’s random numbers were widely distributed and used for solving problems in many fields. RAND researchers were responsible for writing and defending their own reports. The resulting publications, aimed at air force generals, not academic colleagues, were distinguished by a clarity, economy, and self-contained documentation rarely seen in academic work.

  RAND’s first published study was the 324-page Preliminary Design of an Experimental Earth-Circling Spaceship, issued on 2 May 1946. The report advised the government that “the achievement of a satellite craft by the United States would inflame the imagination of mankind, and would probably produce repercussions in the world comparable to the explosion of the atomic bomb. . . . Whose imagination is not fired by the possibility of voyaging out beyond the limits of our earth, traveling to the Moon, to Venus and Mars?”41 When the Soviets launched the first Sputnik, on 4 October 1957, the American public was taken by surprise, but it was no surprise to RAND, whose analysts had predicted the appearance of a Soviet satellite on 17 September 1957, the centenary of the birth of Konstantin Tsiolkovsky, the great Russian rocket pioneer. In November 1957, after Sputnik II had circled the earth with a payload of 1,120 pounds (including the dog Laika), there was no longer any doubt that the Soviets were thinking about launching not only dogs, but bombs as well. U.S. spending on the Atlas ICBM program jumped from $3 million in 1953 to $161 million in 1955 and $1,300 million in 1957. RAND studies led the way. “By 1953 RAND’s knowledge of the missile and weapon fields indicated that nuclear warheads could be carried by rockets, and could produce a wide enough zone of destruction to more than compensate for their aiming errors,” reported a RAND history in 1963.42

  With weapons multiplying on both sides, RAND sought to identify stable nuclear strategies, hosting a renaissance of mathematical game theory beginning where von Neumann and Morgenstern had left off. The titles of some published reports convey the general drift: The noisy duel, one bullet each, arbitrary nonmonotone accuracy (D. H. Blackwell, March 1949); A generalization of the silent duel, two opponents, one bullet each, arbitrary accuracy (M. A. Girshick, August 1949); A loud duel with equal accuracy where each duelist has only a probability of possessing a bullet (M. A. Girshick and D. H. Blackwell, August 1949); The silent duel, one bullet versus two, equal accuracy (L. S. Shapley, September 1950); Noisy duel, one bullet each, with simultaneous fire and unequal worths (I. L. Glicksberg, October 1950).

  The Distant Early Warning (DEW Line) system constructed in 1953 offered a two- or three-hour warning of Soviet bomber attack. By the time the Ballistic Missile Early Warning system came on-line in 1960, the United States was reduced to fifteen or thirty minutes’ warning of an incoming missile attack. Nuclear stability depended on mutual deterrence, either by threatening to launch missiles at the first sign of enemy attack or by preserving the ability to retaliate after a strike. Launch-on-warning, serviceable as a bluff, would be suicidal in practice, since sooner or later an erroneous warning would arise. So the best way to prevent a nuclear nightmare appeared to be to construct a retaliatory system designed to survive attack. To hide, disperse, or harden missiles was comparatively easy; the difficulty was how to construct a robust system of control.

  “Throughout history successful generals make their plans based upon enemy capabilities and not intent,” recalled Paul Baran, an electrical engineer who arrived at RAND at age thirty-three in 1959. “This period was the height of the cold war. Both the U.S. and USSR were building hair-trigger nuclear ballistic missile systems. The early missile-control systems were not physically robust. Thus, there was a dangerous temptation for either party to misunderstand the actions of the other and fire first. . . . If we could wait until after attack rather than having to respond too quickly under pressure then the world would be a more stable place.” Although use of the word surrender was forbidden by act of Congress, RAND officials recognized that “a survivable communications network is needed to stop, as well as to help avoid, a war.”43

  A 1960 RAND study, Cost of a Hardened, Nationwide Buried Cable Network, estimated the cost of providing two hundred facilities with communication links hardened against 100-psi blast pressure at $2.4 billion, with protection to 1,000 psi available for about $1 billion additional cost.44 Baran was assigned the job of analyzing whether it was possible to do as well or better for less. Retaliatory ability could be met by an extremely low bandwidth channel, or “minimal essential communications”—official terminology for the president’s (or his successor’s) ability to issue the order “Fire your missiles” or “Stop.”

  Baran began with Frank Collbohm’s suggestion that by installing a minimal amount of digital logic throughout the existing nationwide network of AM radio stations, a decentralized and highly redundant communication channel could be engaged in the event of an attack. The strategy was to flood the network with a given message. No routing was required other than a procedure that ceased transmission of a specific message once copies of it started bouncing back. Analysis showed that even with widespread destruction of network nodes, with metropolitan broadcast facilities disappearing first, the overlapping nature of the system would allow messages to diffuse rapidly throughout the surviving stations on the net (an unstated conclusion being that American country music could survive even a worst-case Soviet attack). Baran took this proposal to the armed forces, which held out for more bandwidth, believing that real-time voice communication was required to fight a war. “Okay, back to the drawing board,” said Baran, “but this time I’m going to give them so much damn communication capacity they won’t know what in hell to do with it all.”45 And he did.

  Baran’s first job had been with the Eckert-Mauchly Computer Company in 1949. The weaknesses of vacuum-tube-delay-line computers at that time suggested a tenuous future for such unwieldy and temperamental machines. In ten years everything had changed. Witnessing this revolution encouraged Baran to question other assumptions as well. The creative approach to digital computing that had flourished in the early 1950s at the Institute for Advanced Study continued to flourish in the early 1960s at RAND. Baran’s thesis advisor at UCLA was Gerald Estrin, who had been instrumental in disseminating the computer project at the IAS. JOHNNIAC, the RAND version of the IAS machine, incorporated several improvements, including a working Selectron memory, and was completed under the direction of Willis Ware, also an alumnus of the IAS. RAND had steadily acquired the latest machines from IBM. Working under the auspices of the computer department, not the communications department, Baran’s analysis of communication problems was able to develop unencumbered by preconceived ideas. “Computers and communications,” he remarked, “were at that time two totally different fields.”46

  Baran invented a new species of communications network, starting with clear-cut objectives and little else. “An ideal electrical communications system can be defined as one that permits any person or machine to reliably and instantaneously communicate with any combination of other people or machines, anywhere, anytime, and at zero cost,” he wrote. “It should effectively allow the illusion that those in communication with one another are all within the same soundproofed room—and that the door is locked.”47 He threw out all existing assumptions. “In most communication applications, silence is the usual message,” he later explained.48


  By digitizing all communications and multiplexing across the entire network rather than over one channel at a time, Baran knew he could reduce most of the waste. Taking a cue from the store-and-forward torn-tape telegraph networks, he proposed relaying digital messages from node to node across the net, but with high-speed computers instead of telegraph equipment providing switching and storage at the nodes. He examined existing military communications switches and asked: “Why are these things so big? Why do these switching systems require rooms and rooms of stuff?” He concluded that there was far too much recording and storage of messages at the switching nodes, for no apparent reason other than a tradition “to be able to prove that lost traffic was someone else’s fault.”49 Computers were already operating at multimegacycle rates, and Baran knew that the telecommunications infrastructure would eventually be forced to catch up. He proposed a system operating at up to 1.5 million bits per second over low-power, line-of-sight microwave links—ample bandwidth for real-time transmission of digitally encrypted voice. This proposal secured the undivided enthusiasm of the military commanders, if guaranteeing stiff resistance from the managers of AT&T, which handled all military voice communications within the United States. They were not about to admit that their system was physically vulnerable, inefficient, or insecure.

  In May 1960, RAND released the first official memorandum that provided an outline of Baran’s design. “If war does not mean the end of the earth in a black and white manner, then it follows that we should do those things that make the shade of grey as light as possible,” wrote Baran in his introduction to Reliable Digital Communications Systems Utilizing Unreliable Network Repeater Nodes. “We are just beginning to design and lay out designs for the digital data transmission systems of the future . . . systems where computers speak to each other. . . . As there does not seem to be any fundamental technical problem that prohibits the operation of digital communication links at the clock rate of digital computers, the view is taken that it is only a matter of time before such design requirements become hardware . . . where the intelligence required to switch signals to surviving links is at the link nodes and not at one or a few centralized switching centers.”50

  Baran then suggested how to evaluate the robustness of his design: “To better visualize the operation of the network, a hypothetical application is postulated: a congressional communications system where each congressman may vote from his home office. The success of such a network may be evaluated by examining the number of congressmen surviving an attack and comparing such number to the number of congressmen able to communicate with one another and vote via the communications network. Such an example is, of course, farfetched but not completely without utility.”51

  The more alternative connection paths there are between the nodes of a communications net, the more resistant it is to damage from within or without. But there is a combinatorial explosion working the other way: the more you increase the connectivity, the more intelligence and memory is required to route messages efficiently through the net. In a conventional circuit-switched communications network, such as the telephone system, a central switching authority establishes an unbroken connection for every communication, mediating possible conflicts with other connections being made at the same time. “Such a central control node offers a single, very attractive target in the thermonuclear era,” warned Baran.52 The stroke of genius at the heart of Baran’s proposal was to distribute the requisite intelligence and redundancy not only among the individual switching nodes, but also among the messages themselves.

  Baran credited the ancestry of this approach to Theseus, a mechanical mouse constructed by information theorist Claude Shannon in 1950. Guided by the intelligence of seventy-two electromagnetic relays, Theseus was able to find its way around a 5 × 5 maze, in Warren McCulloch’s words, “like a man who knows the town, so he can go from any place to any other place, but doesn’t always remember how he went.”53 To adapt to changes in the layout of the maze and the location of the goal, Theseus had to be able not only to remember but to forget. Baran saw that the problem of routing a mouse through a maze was equivalent to the problem of routing messages through a communications net. “In a very short period of time—within the past decade, the research effort devoted to these ends has developed from analyses of how a mechanical mouse might find his way out of a maze, to suggestions of the design of an all-electronic world-wide communications system,” he wrote in 1964.54

  Baran christened his technique “adaptive message block switching,” abbreviated to “packet switching” in 1966 by Donald Davies, working independently at the U.K. National Physical Laboratory. The first order of business was to take all forms of communicable information—text, data, graphics, voice—and break it up into short strings of bits of uniform length. To the network, all forms of communication would look the same. Error-free transmission of complex messages would be facilitated, for the same reason that the reproduction of complex organisms can best be achieved in a noisy environment through the collective reproduction of large numbers of smaller component parts. Every bit of data that passes through a network acquires a certain probability of error, and the cumulative probability of one of these errors affecting any given message grows exponentially with the message length. Short segments of code are far more likely to arrive unscathed. Error-detecting processes are most economically applied to short strings of code, checking each message segment for errors and only retransmitting those individual segments that fail. This is analogous to proofreading a manuscript one page at a time and retyping only those pages containing mistakes. Baran proposed using cyclical redundancy checking, or CRC, a method widely utilized today.

  Each 1,024-bit message block was flagged to distinguish it from its neighbors and provided with fields containing its “to” and “from” address, as well as information needed to reconstruct the original message sequence at the other end. The message block also included a handover tag, which was incremented every time the code segment passed through a node. Each individual code packet knew where it was going, where it had come from, which message it belonged to, and how many steps it had taken along the way. This information was shared with the host computer every time a packet passed through a node. “Packets are best viewed as wholly contained information entities designed to match the characteristics of the network switches,” noted Baran.55

  At the nodes, simple procedures (dubbed the “hot potato heuristic routing doctrine” by Baran) kept packets moving in the right direction and ensured that the network would adjust itself to any congestion or damage that arose. By looking at the “from” address and the value of the handover tag, the station kept track of which links were forwarding messages most efficiently from any given address and used that information to direct outgoing messages, operating on the assumption that the best link coming in was likely to be the best link going out. “Each node will attempt to get rid of its messages by choosing alternate routes if its preferred route is busy or destroyed. Each message is regarded as a ‘hot potato,’ and rather than hold the ‘hot potato,’ the node tosses the message to its neighbor, who will now try to get rid of the message.”56

  Individual switching nodes learned from experience where each station was and quickly updated this knowledge if a given station was damaged or moved. Baran found it possible to “suggest the appearance of an adaptive system” by implementing a “self-learning policy at each node so that overall traffic is effectively routed in a changing environment—without need for a central and possibly vulnerable control.”57 A demonstration system, simulated as a forty-nine-node array on RAND’s IBM 7090 computer, proved to be surprisingly robust against both random failure and deliberate attack. Starting from a “worst-case starting condition where no station knew the location of any other station,” it was found that “within ½ second of simulated real world time, the network had learned the locations of all connected stations and was routing traffic in an efficient manner. The mean measured path length compare
d very favorably to the absolute shortest possible path.”58

  The complete study was published in a series of eleven reports released in August 1964. An additional two volumes on cryptographic issues remained classified, even though, as Baran explained in the ninth volume, devoted to security and secrecy, “if one cannot safely describe a proposed system in the unclassified literature, then, by definition, it is not sufficiently secure to be used with confidence.”59 Baran later emphasized that “we chose not to classify this work and also chose not to patent the work. We felt that it properly belonged in the public domain. Not only would the US be safer with a survivable command and control system, the US would be even safer if the USSR also had a survivable command and control system as well!”60

  In late 1962 it was estimated that each switching node would occupy seventy-two cubic feet, consume fifteen hundred watts of power, weigh twenty-four hundred pounds, and incorporate 4,096 32-bit words of memory. By 1964 the estimate was down to units occupying one-third of a cubic foot, with twice the memory, no air-conditioning, and costing fifty thousand dollars less. Unlike most military budgets, this one was going down. By the time the final reports were released, most skeptics had been persuaded of the project’s advantages—“the only major outpost of frontal opposition was AT&T.”61 In 1965, RAND gave the air force its final recommendation to implement the project in a preliminary configuration comprising two hundred multiplexing stations distributed across the continental United States, each capable of serving up to 866 simultaneous subscribers with data and 128 with voice. There were to be four hundred switching nodes, each supporting up to eight full-duplex “minicost” microwave links. By 1966 the air force had received a favorable evaluation from an independent review. Only then was it determined that jurisdiction over the project would have to be assigned not to the air force and its independent contractors but to the Defense Communications Agency, which Baran and some influential colleagues believed was ill prepared to construct a nationwide network based on digital principles at odds with the communications establishment of the time. Baran was forced to advise against the implementation of his own project, lest “detractors would have proof that it couldn’t be done . . . a hard decision, but I think it was the right one.”62