THE MN-MACHINE RELATION IN METEOROLOGICAL DATA PROCESSING by ROBERT CHARLES GAI4MILL B.S., University of Rochester (1959) SUBMITTED IN PARTIAL FULFILLMENT OF THE REQUIREMENTS FOR THE DEGREE OF MASTER OF SCIENCE at the MASSACHUSETTS INSTITUTE OF TECHNOLOGY June, 1963 Signature of Author Certified by Accepted by ..... ....... .... ....... ..... ,................. Department of Meteorology, May 17, 1963 ..... .. .. ...... .......... ............. s.........p ... Thesis Supervisor . * . ....... Chairman, Departmental Committee on Graduate Students ABSTRACT THE MAN-MACHINE RIATION IN METEOROLOGICAL DATA PROCESSING by Robert Charles Gammill Submitted to the Department of Meteorology on May 17, 1963, in partial fulfillment of the requirement for the degree of Master of Science At present operational weather forecasting on a com-puter is an incompletely programmable problem. This situ-ation has every prospect of continuing to be so. Such problems require the application of intelligence in the attempt at solution of their unprogrammable aspects. We have two alternatives in providing this: use of 1) human intelligence 2) artificial (machine) intelligence Artificial intelligence is not yet adequate for our task, and may not be for many years. Thus, we must find means for direct application of human intelligence in our problem solution. Man-machine symbiosis is the method by whtoh we can achieve this. Symbiosis is the complementary utilization of the talents of man and machine in a system which is more capable than both workinG separately. That is, each works on those portions of the problem at which he is most capable. A possible future weather central, using such a philosophy, is described. Feasibility of on-line human manipulation of isopleth patterns is demonstrated. Thesis Supervisor: frederiok 'Sanders Title: Associate Professor of Meteorology 3 TABLE OF CONTENTS 1. Operational Weather Forecasting 1.1 Data Collection 1.2 Data Processing 1.3 Useage 2. The Weather Central Today 2.1 Equipment 2.2 Programming 3. Man Computer Symbiosis 3.1 Problem Areas for Man-*Computer Symbiosis 4. Feasibility of Meteorological Symbiosis 4.1 The Computer 4.2 The Program 4.3 Results 5. The Weather Central of the Future 5.1 Equipment 5.2 Programming FIGURES Number Page 14 2 3 4 9 10 19 5 36 6 7 8 9 10 11 12 13 38 38 39 39 40 40 41 41 1. OPERATI ONAL WEATHER FOR It OPERATIONAL WEATHER FORECASTING This thesis will consider the problem of the most effective design for a weather central Using automatic data processing techniques. A weather central is an integral part of the data collection, processing and distribution system which provides operational weather forecasting ser. vice to users here and abroad. processing phase. Its primary job is the In order to reach concrete conclusions about the design of a central, we must examine the job to be done, and the materials to be used. Operational weather forecasting has been going on for many years. It started long before anyone had a clear idea of the physical mechanisms of the atmosphere. Today, with more knowledge about the mechanisms, we are in a better position, but the processes through which forecasting services are provided are essentially the same. A simplified flow diagram of what occurs is shown below: a) DATA COLLECTION b) PROCESSING F1 c) USE - - data sample processing -picture of the spacem - time character f igure 1 I-... ..... ..- - _ 5 a) Data gathering must be treated as a sampling problem, for the size of the population (the molecules of the atmosphere) makes complete measurement of its characteristics unfeasible. b) The processing system (meteorologist or computer) converts the data sample into a space-time characterization of regions of the atmosphere* a) This characterization is given to the user, who uses the information in making decisions. The purpose of the system is to provide the user with knowledge of the future (and occasionally past or present) location or distribution of meteorological phenomena which affect him. The primary concern of this paper will be the design of a processing system. As will be seen later, processing is the only phase over which it is possible to exert much control. The great majority of data collection agencies are uncontrollable due to interagency and international boundaries. Thus, the safest approach is to design a processing system which can adapt to existing conditions of data input, and hope for the best in the improvement of data collection methods. Since we will be assuming that the present cond- itions of data collection are unalterable, it is important to consider what those conditions area 1.1 DATA COLLECTION There are two major types of meteorological observations. These are: 1) radiosonde 2) surface. The radiosonde observation is made from a helium filled balloon, transmitting information about temperature, pressure, and relative humidity at certain levels as the balloon rises. Wind speed and direction at desired levels can also be found by tracking the balloon using radar or theodolite. These observations are taken at least every 12 hours (at 0000 and 1200 Greenwich Mean Time) by approximately 500 Northern Hemisphere stations. Because of the number of instruments which are thrown away under this system, they are inexpensive and somewhat error prone (the radiosonde is not one of the weakest links in the data collection system). Furthermore, the data which reaches the Ground must undergo considerable computation before it .s prepared for transmission. Thus, there are at least two ways that a radiosonde observation may be in error before it leaves the observing station. In addition, the most important observations are those taken under difficult conditions, such as during a thunderstorm, from shipboard, or from an isolated station. Also, different countries use different instruments, different units, and different procedures, with the result that the data will be subject to error of many types. The standardization and improvement of this data collection method is a slow and continuing process. part in such change. Political considerations play a large Surface observations are subject to a different type of problem than the radiosonde. Since the instruments are not discarded after each use, and since there is little computation to be done, surface observations tend to be more accurate. However, because of the ease with which they may be taken, and the need for heavy coverage, many are taken by people with little meteorological experience and sometimes little interest. Especially in ship observations, errors are found due to incorrect use of the instruments, difficulties of shipboard measurement, and carelessness. Since these ship observations are supplied on a voluntary basis,.little can be done to improve their quality. Because they are taken in regions of sparse coverage, they take on great importance. Many other types of observations are made, the most important being pilot reports from aircraft. Here, instru- ment inaccuracies, errors of observation location, carelessness, political problems, and other factors, limit our ability to improve observation accuracy. These observations also tend to be more important because they often come from areas of sparse coverage. Grave difficulties stand in the way of hopes for substantial improvement or standardization of global observational data. Even the difficulty of achieving further standard- ization and accuracy within the United States should not be blithely discarded, for politics like crabgrass is ubiquitous. 8 Moving to the next phase of data collection, once the observation and associated calculations have been completed, the data are ready to be encoded into a report. The codes which are used are catalogued in handbooks distributed by the World M-eteorological Organization. Variations in the coding format occur due to observation type, time, national practice, agency (Air Force, Navy, FAA, Weather Bureau to cite a few), and sometimes individual preference. Some of these variations are catalogued in the handbooks, many are not. Major changes in format sometimes occur with little notice. After the report has been encoded it is transmitted by teletype, radioteletype or radio. Before the report reaches our processinS center it may have been retransmitted up to ten times, come halfway around the world, been edited one or more times, had its units converted, been decoded and recoded in another format (summaries eto4, and run throuGh several storage devices, such as paper tape punches or magnetio tape recorders. Almost all the processes described involve human activity. At every one of the points described, human, electronio, or mechanical error may be introduced. A further problem inherent in the data-gathering phase is the timing problem, or coverage, Because of the meander-* ing path which a report may take in reaching us, the amount of human intervention necessary, and the numerous poesibilities for communications problems (electronic disturbances 9 or equipment failure), many reports will not be available in time to be included in our data sample, or will be so garbled that they will be unusable. This problem Is not serious in this country, but in the South Pacific, where regions the size of the United States have no data, one observation can mean the difference between a reasonable forecast and nonsense. Further, the regions of sparse data are those in which communication ability is least assured. Reports from sparse data areas are often difficult to obtain, and to interpret once received. Yet, they are extremely important since the atmospheric mass which they must be used to characterize is so enormous. In summary, the data acquisition problem may be flow diagrammed as below: figure 2 instrument . and human errors ATMOSPHERE instrument q human inst ent linstrument human I 1human 1. transmitter transmitter Iraceiver. .receaver Itransmitter ... .. transmission line losses instrument and human errors transmission line losses instrument errors ZeceiverI. human transmitter tran mitt trsmit receiver receiver r PROCESSING STATION i * 10 It must be reiterated that the imperfection of the data collection network is a much stronger function of politics than of technology. However, even if the political problems were to be miraculously dispersed, those observation points most desperately in need of sophisticated measurement and transmission technology are those least able to afford and maintain the necessary equipment. 1.2 DATA PROCESSING The data processing phase be3ins with the reoeipt of the data. Immediately we are faced with the problem of determining how long after observation time (JMT at which the observations were taken) one must wait before the data sample can be considered complete. We want to find that point in time where a further wait will not increase the accuracy of the output sufficiently to overcome its decreased usefulness due to lateness& data time ~- sum of both completeness of samplle timeliness figure 3 time Once a decision has been made, our limitations in time become well defined. Next in order is the data identification, decoding, consistency checking, and interpretation. The different parts of the processing phase are strongly interrelated, but will be separated for clarity. Identification of reports takes two forms 1) station identification 2) classification of report type Stations are identified by a five digit number, the first two digits representing the block number (a geographic and politioal grouping of stations) and the final three digits the station number within that block. Reports not meant for international distribution (e.g. from stations in North America) usually do not include a block number, so that ambiguities sometimes arise. The identification process involves the supplying of a location for the report, either from memory, by use of a map of station locations, or use of a library of station numbers and locations (in computer methods). Extreme difficulty is encountered when the station identification is incorrectly received or when no location is known for the particular identification. In these cases fairly sophisticated detective work is necessary. Classification of report type is not always easily accomplished. Report lengths vary from five to 100 words (each word approximately five characters). There are ten to twenty major types of report, each having many variations within its own format. Some reports use alphabetic and special symbols, others use only numbers. In some reports all the words have exactly five characters; in others various word lengths occur. Some of the information trans- mitted over meteoroloGical teletype lines is written in plain language, except that it is abbreviated, and the set of abbreviations to be used is the prerogative of the sender. Some report types strongly resemble another, but incorrect classification will give meaningless information. Within a bulletin (group of reports identified by a heading) there may be several types of reports. Thuso it is not possible to set down a simple set of rules for report classification. Two types of consistency may be checked for: 1) format (coding structure) 2) physical (satisfaction of physical constraints on a fluid) Format checking is the use of redundancy of report structure to discover and, when possible, correct garbling. This is normally done during the decoding of the report, le., the conversion of the coded information to meaningful meteorolog-ical parameters. may beGin. Once this is accomplished, interpretation During interpretation our knowledge of the physical constraints on the atmospheric fluid is used to indicate parameters which are doubtful. Interpretation takes two forms: 1) analysis (two dimensional mapping on constant pressure surfaces) 2) vertical analysis on aerological chart (one dimensional mapping of individual radiosonde report) Analysis is the plotting on a map of all the data which is 13 available at a pressure level and the addition of isopleths (lines of constant value of the quantity) to indicate the distribution of the parameter of interest. In computer methods the values at a predetermined grid mesh are obtained by interpolation from the observations. Both contouring and grid mesh production are essentially a two dimensional integration of discrete measurements of a surface, since the family of isopleths may be regarded as depicting the topo-W graphy of a surface. Our filtering process is one of looking for unrealistic space variations of the present surface and of the time change of the surface since the last data time. Analysis can be of great use in detecting incorrect data, for much is known about constraints on horizontal patterns, both spatially and in time, and about the meanings of significant features of such patterns* A problem concerning our know.' ledge of atmospheric constraints is its conditional character. For example, wind speeds of 150 mph are rare, except in the, vicinity of the jet stream, or of a hurricane. Thus, normal constraints on the atmosphere break down just when we are most interested in accuracy. We are most likely to disbelieve a report which indicates a strong disturbance, and believe one that shows nothing of great interest. Vertical analysis is the plotting of reported data from a radiosonde Ieport on an aerological chart. Much can be learned from the vertical distribution of temperature,. 14 dew point, and wind velocity. Constraints are also well defined in the vertical, especially with relation to lapse rates (the rate of deorease of temperature with pressure). There is a limiting lapse rate that is likely to occur over any but extremely shallow atmospheric layers. Here again, we will have a tendency to disbelieve indications of unlikely occurrences. If true, these are of considerable interest. Once a three-dimensional, present-time picture of the atmosphere has been produced through the two means described, the process of forecasting begins. This process takes many forms, depending upon the type of forecast desired and the time range for which it must apply. Forecasting is the extrapolation of our three-dimensional picture of the atmosphere to a future time through an Integration process. The forecasting process is highly complex, and much effort has been devoted to development of forecasting models in recent years. Forecasting schemes which are presently used are of, two types: experience, ie. statistical, and dynamical. Statistical forecasting is a recent development associated with the systemization of experience forecasting, Experience forecasting is based on concepts such as: 1) Persistence Statistically the most likely future occure rence, for short time periods, is no change. 2) EXtra olation Weather systems tend to move at a velocity proportional to the wind velocity at 500 millibars (about 18,000 feet). 15 Statistical forecasting makes use of the past records of the time behavior of weather systems to predict the behavior of similar present systems. Statistics also provide us with a means for predicting the most likely weather for a time period, known as olimatology. Dynamical forecasting is one of the more recently developed areas of meteorology. The advent of large digital computers has made possible the approximate solution of the non-linear equations of fluid dynamics, specialized for the large-scale motions of the atmosphere, in an amount of time which makes forecasting by this method feasible. It should be noted that these techniques have found direct utility only in wind forecasting, nevertheless, dynamical forecastinG has had a Great impact on meteorology* At present there are at least two weather centrals in the United States making suoh forecasts. All is not sweetness and light in the field of dynamical forecasting, however. Due to causes such as the addition of scale assumptions (to prevent energy from appearing in very short iwave lengths and masking meteorologically interesting wave lengths), the equations represent the atmosphere incompletely. Thus, although dynamical pre. diction schemes can exceed purely human capabilities on the average, there are many situations (statistically infrequent but often important) in which a man is able to recognize that the dynamical saheme is in error. 1*3 USEAGE The users of processed products are usually not intere* ested in large geographio areas. They are interested in a localized reGion or, in the case of aircraft, a route. The individual user Is interested in a variety of parameters that is large, and not necessarily known. Thus, in most cases it is not possible to tailor output to the users9 needs; instead, products desoribing the overall atmospheric character are sent to him to be interpreted in terms of their impact on his area of interest. This interpretation usually involves examination of the present situation, and of the products received. If the products do not show significant activity, the forecast will usually be for persistence of present conditions plus any statistically known variations due to local characteristics. Tailoring of output to user needs is done occasionally, but is difficult because requirements often change. An example of such tailoring occurs in the Global Weather Central at SAC headquarters where route wind forecasts are made for particular aircraft routes. Considerable difficulty can occur when routes are changed rapidly, or when require4 ments for special accuracy arise. 17 2. THE WEATHER CENTRAL TODAY The weather centrals described in the following section are the National Meteorologioal Center at Suitland, Maryland and the Global Weather Central at Strategic Air Command Headquarters, Omaha, Nebraska. Centralized processing of weather data existed before the development of automatic data processing systems, but the volumes of data involved in large scale weather prediction make the use of a computer an obvious step. The- development of dynamical prediction models and sophisticated statistical methods has made the use of computers mandatory, for the associated computation is prodigious. 2.1 EQUIPMENT Both centrals use the IBM 7090 as the main processor (with a probability of conversion to 7094 in the future). They both have numerous magnetic tape units, an on-line printer, and an on-line card reader. Both have recently added an IBM 1301 disk file, for large--volume rapid-access memory. At both centrals off-line output is handled in two ways: 1) IBM 1401 with magnetic tape units, paper tape punch, and high-speed line printer. 2) Electronics Associates Data-Plotter with magnetic tape unit. The output devices listed are used for display of final output 'in the form of contour maps or messages for teletype transmission. Both of these methods produce hard copy which is then displayed for reference, transmitted over facsimile or teletype circuits, or filed. Display of intermediate results for monitoring purposes must be on the on-line printer, which is very slow, so monitoring must be kept to a minimum. In their approaoh to input, the two centrals differ slightly. NMC uses special purpose gear designed by AVCO, which puts the teletype messages on maGnetic tape to be processed by the 7090. GWC has been using its 1401 in the conversion of teletype paper tape to magnetic tape. Under development is a processins system in which the ADX computer of International Telephone and Telegraph is used for realtime processing of teletype data direct from transmission lines. This computer will do some processing of the input data and place it on the disk storage to be used by the 7090. It is probable that direct output to teletype lines will be handled by this computer in the future. 2.2 PROGRAMMING Programming for automatic weather data processing systems has grown as an adjunct to dynamical prediction. Initial applications of computers to meteorolo3y were only in that area. As attempts to make dynamical prediction operational have borne fruit, it has become necessary to produce programs quickly to automate the input and output phases of processing, In trying to produce such programs it has been found, particularly with respect to input, that there are some rather difficult problems. In general, the following, flow diagram describes the programs necessary for numerical weather predictions FLOW DIAGRAM 0F' AUTOMATIC WEATHER DATA PROCESSING SYSTEM INPUT DEVI CE CHECKING OUTPUT identification station in library? report type and location decode format consistency standard format report hydrootatic (vertical) analysis vertical physical consistency vertically consistent report analysis horizontal and time consistency srid maps for present time at levels: sfc., 1000 mb., 850 mb., 700 mb., .. forecast grid maps for future times at levelas sfc., 1000 mb., 850 mb., 700 mb., ... verification statistical ability of forecast method output programs_ contour maps and teletype messages DISPLAY OR TRANSMISSION DEVICE figure 4 20 Programming is of the nowntop variety due to the hish cost of computer time and larGe volumes of data to be processed. Programs in sequence are treated as entities, each taking its input from magnetic tape and putting its output on another tape. Each of the proj;rams has no communication with the other programs except through input and output. From the fore3oion discussion there is considerable doubt whether a system of the type presently used can ever do a completely satisfactory job of operational weather data processing. Some decisions must be made in a rather arbi- trary manner, and most algorithms incompletely represent the methods used by a human being. The result is that difficulties arise in the following areas: 1) station identification and location 2) classification of report type 3) deearbling through format consistency 4) standardization of units and correction of systematic errors 5) ooverage sufficiency or requests for missing data 6) reasonableness of vertical patterns 7) reasonableness of horizontal and time patterns 8) reasonableness of forecast patterns 9) adaptability for special output Because of the numerous problem areas, and the lack of human control, products of automated data processing facilities are sometimes unsatisfactory and almost always less accurate than they miGht be with human assistance. For this reason automatic weather data processinG centers have served pria marily in a supplementary capacity. They have not sup- planted the large number of human beings in the weather central. Some of the former manipulative labor has been takoen over, but, for the most part, humans are doing the same jobs they have always done. Why has the computer been unable to replace the human? There are two reasons: 1) the need for reliability 2) the necessity for adaptive or intelligent behavior The need for ieliability is most easily illustrated by the situation at the Global Weather Central. The strategic offensive capability of the United States is partially dependent upon the availability of high quality products from that facility. Thus, to Guarantee reliability in spite of equipment malfunction or program failure, it is necessary to supply sufficient personnel to perform, by hand, operations similar to those performed by the computer. Such a means of providing back-up capability is extremely costly. Equipment reliability is achievable through means which have become well developed in the space and commandcontrol industries. In particular, the use of redundant components can lead to extremely high reliability. An example of the use of redundancy in a successful computer system is SAGE of the Air Defense Command. Such a system gives extremely high assurance of continuous operation, alone 22 with the high precision we have come to expect from the use of computers. Program reliability is somewhat more difficult to achieve. Here one finds the problem of large numbere of eventualities and the necessity for maintaining high quality in spite of unforseen circumstances. Such a problem requires the ability to adapt, or apply intelligence. us to the second area of concern. This brings The study of artificial (machine) intelligence is progressing, by leaps and bounds, but is still not equipped to handle problems of the real world. Progress is being made in the solution of problems which have well defined rules, 1e Games esuch as chess and checkers, and mathematical theorem proving. However, the level of intelligence being exhibited by machines is still too low to be of use in weather data processinG. Our answer lies not in programming intelligence into the computer, but in man-wcomputer cooperation (or symbiosis, as it has been termed by Licklider, 1960). 23 3, MAN-wCOMPUTER SYMBIOSIS Man-computer symbiosis is a subclass of man-mmaohine systems. Certain types of man-machine systems have been in existence for many years. Examples of such mechanisms are the bicycle, the adding machine, or the television set. allows man to extend his capabilities, calculating ability or his eyes. Each .e* his legs, his However, in each of these mechanisms, man provides the initiative, the driving force, or the direction. These systems do not constitute symbiosis, but the mechanical extension of man. In modern computer centered data processing systems we find a drastic change from the mechanical extension of man. Here we find that computers have taken over the greater portion of the task, and are being aided by humans. This has been termed the human extension of machines. It is the other end of the man-machine interaction spectrum* Any humans present are there to help the machine through maintenance or the performance of tasks which have proven difficult to automate. Such systems are not symbiotic. which started ou They are systems to be fully automatic, but didn't quite make it. This is the present state of the computer centered weather central. An extension of the spectrum, presently under develop-t ment, is artificial intelligence. The development of intell- igence in the computer should remove the need for human aid, and allow the computer to forge ahead on its own, Such a development may well invalidate the need for symbiosis; however, the age of artificial Intelligence is not yet here, and it is anyone's guess when or if it will arrive. It has been estimated that it will be 1980 before machines alone will be able to do muoh useful thinktIng on military problems. Licklider (1960) has allowed the possibility of up to 500 years. Another point in the man-machine spectrum is man-* computer symbiosis. This is the most advanced means avail-. able at present for solution of problems requiring the exerelse of intelligenoe. It is hoped that through the close coupling of complementary talents of man and computer many tasks which cannot be handled satisfactorily by either alone may be approached through cooperative effort. Symbiosis is an important development in methods of computer useaGe, for the tendency in recent times has been to a continual decrease in man-computer couplinG. Because of the high costs of computer operation and volumes of programs to be run, most computer facilities have instituted the policy of disoouragin- pro-rammer presence durin2 run. Instead, pro. grams are submitted with instructions for operation, and run in large batches under an automatic system. hours later the results are returned. Several This means that work-a in.,* with the computer in producinG a problem solution is a slow and tedious process. Thus, one way in which manwcompu-& ter coupling may be judged is by the time it takes from test or query submission until results or a reply are fortheoming. 25 Another area in which man-computer coupling may be examined is the ability of each to communicate complex ideas to the other. The goal of symbiosis is to reduce reply time, increase the ability of man and computer to communicate, and, as a result, make feasible the close association of man and computer in problem solution. It is of interest to examine the types of problems around which present day computation systems are oriented, and those which are unfeasible or unattractive as a result. Present systems require problems in which all eventualities can be forseen and all procedures preplanned. For many purposes this is no disadvantage, but it becomes important for certain classes of problems, those which are difficult or impossible to formulate or preplan. Meteorological tasks are of this sort. One type of problem that is difficult to preplan is one in which a large set of possible occurrences require a large set of reactions* Although one may program for a great number of eventualities, some will be ignored. Receipt of unsatisfactory results and corrective programmin3 is a conm tinual process. ProGrams requirinG this approach are common in meteorological data processing. If the set of unprogram- med eventualities is large and the time required for one test is also large, improvement becomes very slow. Symbiosis, by cutting down test time, allows us to formulate our problem solution initially in a crude manner and rapidly develop 26 better procedures through trial and error. Symbiosis can also help us in another way. Once a program of the sort described has been developed to the point where its procedures are fairly adequate, it may be used operationally. Although many eventualities will not have been found, through our close cooperation with the computer we can monitor (quality control) its actions and take manual corrective action if the need arises. The information gained can be used later to improve the proGram. If the problem is completely programmable (forseeable or formulatm able) this second phase will become unnecessary. From all indications most meteorological tasks will require such monitoring, Another type of problem which is difficult or impossible to handle with present data processing methods are those which include eventualities or actions which are too oomplex to be easily described. Here, the monitoring aspect of symbiosis is the most important, for we find that we are completely unable to formulate the problem in machine terms. In such problems the human is able to fill the cap. An example of this type of problem is one which includes reoogenition of complex patterns, i. hurricanes, frontal systems, air mass types.etc. In summary, symbiosis is important for problems whose method of solution is difficult or impossible to formulate. Solutions to such problems are usually an approximation to a "best" solution. Thus, we can look on the process of 27 formulation as one of hill climbing- or maximization of ability. Short test time allows one to approach the top more rapidly. The closeness with which a best solution can be approached is determined by the undiscovered aspects of the problem. In any case, an only partially satisfactory solu- tion may be used since a human may now fill in the gaps in its methods. In the gap filling process it may be possible to learn ways to improve the program. 3.1 PROBLEM AREAS FOR MAN-COMPUTER SJBOSIS Now that we have looked at the many reasons why mancomputer symbiosis is desirable, the obvious question arises: why hasn't it been done? The answer is to be found in the differences between man and computer. Some of the same 4ifferences which make symbiosis attractive are those which make its realization difficult. The first difficulty to be examined is that of speed. Everyone knows that computers are fast, and Gettingr faster4 With speed comes expense. The result is that it can be a very expensive proposition to make a large digital computer wait for a human to take some action. This is pre- cisely why most present data prooessing systems do not allow stops and why loading and input-output functions are done without human intervention. For many purposes, the lack of human intervention is a good idea, but how can we economically allow such intervention when it is necessary? The 28 answer is time sharing. This new approach to computer use is presently in operation, and being developed further, at M.I.T. Time sharing allows many users to share the available computer time in an efficient manner. It is essentially an automatic time allocation program which distributes run time to other programs, under a priority system. Through such an approach it is possible to allow any one program to pause and wait for human help, and still be economical, for now another program can be utilizing the time. In other words, instead of one program having complete control of the computer, a pool of programs to be run are available to a time allocation program. inactive programs. The pool is divided up into active and Inactive programs are those which are waiting for human help or for an input-output device to complete some job. time. These cannot immediately use computation Active programs are those which can immediately util- ize computation facilities. Active programs are run for short periods of time in a round robin fashion, under a priority system. The priority system is usually such that man-computer interaction programs receive first attention. Thus, computer reaction to human action is rapid. Programs not requiring man-computer interaction are run at a lower priority, filling in when all other programs are waiting. As can be seen from the above description, time sharing makes mandatory the availability of large-volume, high-speed 29 storage facilities for keeping the pool of programs. In the past this was a difficult requirement to satisfy, but the use of the IBM 1301 disk file (available at both centrals) fulfills this need. There are other problems in making time sharing an efficient process, but it is interesting to note that many of the characteristics of operational weather data processing simplify time sharing. For example, the fact that. the programs being time-shared are operational and, as a result, fairly predictable as to size and actions, can reduce difficulties. In the time sharing systems being developed at M.I.T., it, is assumed that the programs being run are unknown and untested. Furthermore, in an operational set of programs the use of input-output devices may be preplanned, whereas, in a general time sharing system the input-output devices must be allocated by the system. Thus, because of its operational character, the problem of providing time sharing for an operational weather data processing system should be a reduced problem. The second major difficulty in the realization of symbiosis is provision of the ability for man and computer to communicate effectively. This problem has two phases: 1) input-output devices 2) flexible man-computer language The problem of providing hardware for a man-computer communication link is one which has received considerable study by those developing command-control consoles. Available ~ m I - devicea, although adequate to handle the job, are not yet in a really desirable form. At present the best available means for communication is the use of a cathode ray tube (abbreviated CRT) with light-pen and typewriter. More advan- tageous would be a system which allowed verbal instead of typewritten communication, and had a surface for visual display of a higher capacity than the present cathode ray tubes. Presently available equipment can do the job, however, with some limitations. It is important to mention at this point that to be useful the CRT display must be self-maintaining and dynamic. If the computer must be tied up by the display process, then it cannot be doing useful labor, and the result is an uneconomical use of the machine and no need for time sharing. The display must be dynamic so that light- pen techniques may be used. A further desirable feature for the CRT device is the ability to display curved and straight line segments, as well as points. It is possible to do num- erous weather processing tasks with devices which lack some of the capabilities mentioned above (see section 4 of this thesis). The second phase of the man-computer communication problem is the language. Here again, the fact that we are working with a limited set of problems gives us an advantage. The symbol system used in operational weather forecasting is well defined, because, since the beginning of weather service it has been necessary to transmit symbolic information A 31 about~Weather over mechanical devices. As a result, meteor- ologists are well trained in the use of the set of symbols. Furthermore, the kinds of operations necessary for increasing the accuracy of operational weather forecasting do not seem to present a formidable obstacle. Ultimately, it may be possible to allow extremely sophisticated man-computer symbiosis in meteorology through the development of advanced languages, but this is unnecessary for the purposes being described here. ~--~ ~ 32 4. FES4BILITY OF METEOROLOGICAL SYMBIOSIS IA.attempting to demonstrate the ease with which a symbiotic use of the computer may be achievved in meteorology, it was decided to examine the problem of isopleth display and manipulation as a means for control of grid data arrays. Cathode ray tube display devices have been utilized for the display of well defined symbols (alphabetic, numeric, etc.) and graphs for some time. Thus, it was felt unnecessary to investigate the manipulation of coded reports (as they appear on teletype) or plotting of aerological diagrams, since similarity to present uses promises easy development of symbiosis. Considerable work has been done recently in the use of the computer and cathode ray display in computer aided design at M.I.T. Work by Sutherland (including a movie) gives an exciting look at the potentialities of such systems in the manipulation of geometric figures and diagrams. However, Sutherland's work involved the use of circles and straight lines, but no arbitrary curves such as are found in meteorological isopleth patterns. Although there was no hope of achieving equivalent flexibility in the present project, it was felt that an investigation of the manipulation of curves would be a fruitful addition to Sutherland's contribution. The problem, as initially conceived, was the development of sohisticated means for modification and control of grid data fields. These are one of the most common forms of data used in computer weather data processing. Isopleth modification is not the only approach to this problem, nor is it the simplest. One simple way of achieving success at grid field modification would be to allow the choosing of a subset of grid points through the use of the light-pen, and then to cause a constant quantity to be added or subtracted from the field values at those points. Another simple approach would be to allow the addition to the field of a positive or negative point vortex at a position defined by the light-pen. Another method, suggested by normal analysis techniques, is the creation of a hypothetical station report to correct a grid region. Isopleth modification seems to offer the most effective means of grid data control. The isopleth pattern is, in a sense, a summarization of the character of the field. If we could modify the grid field through redrawing of the isopleths we would have flexible control. Another argument for the direct modification of isopleth patterns is that such patterns are the primary symbolic method through which a two dimensional mapping of a quantity can be represented to the human mind. Thus, changes necessary to isopleth patterns, interpreted as movement of position, tend to be more obvious to the human than quantitative changes of the grid values or vorticity. 4.1 THE COMPUTER Problems were encountered in trying to choose a computer. Time sharing is available only on the Computation Center 34 7090 which does not have a light-pen associated with its cathode ray display. Further, no computer at M.I.T. has a The CRT on the TX-2 CRT with self maintaining display. computer at Lincoln Laboratories (used by Sutherland) does. Thus, since the program to be written was obviously going to waste time, it was decided to program it on the TX-0 (a small computer available for student use), which has the fastest display time (approximately 40 microseconds per point) of any computer at the Institute. 4.2 THE PROGRAM The program starts with a seventeen by seventeen grid point array, and produces contour lines for the resulting 256 grid areas through a linear interpolation method. The CRT display device on the TX-0 displays discrete points only, so it is necessary to display curves as collections of points. Not all the points available on the tube face (5122) are used, due to limitations of display speed. Instead, sixteen points (four by four) per area for the 256 (sixteen by sixteen) grid areas are used. If too many points are displayed, the amount of flicker becomes irritating. The locations (on the tube face) of points to be displayed are stored in two ways. Display storage is a table. holding one tube face position per word. Each word uses the left nine bits to represent the X coordinate and the right nine bits to represent the Y coordinate, just filling the 35 eighteen bit word. In this configuration the locations are ready for immediate display by simply loading them into the accumulator and giving a display instruction. However, this storage configuration makes magnification changes difficult. Thus, a codeword storage is used. The codeword store uses one word per grid area, sixteen bits of the word representing each of the sixteen possible point locations in the grid area. This allows very compact (256 words) picture storage, which is unaffected by magnification or centering changes. To create point positions for display, the codeword storage is unpacked. Using magnification and centering information from the switches, a tube face location for each point is computed. for efficiency and speed. These are stored in display storage Each time that the magnification or centering settings are changed, the display storage must be completely erased, and the new point locations computed using the codeword store and the new settings. Display modification is accomplished through use of the light-pen. The light-pen is a photocell which sets a relay when it sees a point displayed. Thus, after each point is displayed, a test must be made to see if the light-pen has reacted. Tracking the light-pen is achieved by the plotting of vertical and horizontal lines of points in four directions from the last known position of the pen. The average of the four first unseen point locations in the four directions is the new pen position. 36 -not seen horizontal average old pen position not seen---+ 0 0 0 + 0 *--not seen -vertical average %-area "seen" by light-pen new pen position * i-eot seen figure 5 Tracking the pen allows us to draw curves on the tube face. Seeing a point which makes up a displayed figure can be used to erase that point, or to find the position of the pen to initialize tracking. Another method of finding pen position is the display of points in random positions over the tube face until one is seen. The program has two modes of operation, drawing and erasing. In the drawing mode (set by a swit-ch) the computer displays the full contents of the display storage watching for a reaction from the pen. If another switch is on, 100 randomly located points are displayed after each pass through display storage, also looking for the pen. found, tracking begins. When the pen is The tracking cross must be displayed often, so only eighteen points from display storage may be shown before the cross must be plotted again. To produce a set of points which represent the track, the program remembers the first location where the pen was seen. As each new 37 pen position is computed, this remembered position is displayed. If it is still in view, the new pen position is used only for tracking. If the remembered location is no longer visible, the new pen position replaces it, and the new position is stored in both display and codeword storage. The other mode of operation is the erase mode. Display storage is again plotted on the tube, but now, when the pen reacts, the position is erased from display storage. The codeword storage representation of the point is erased by removing a bit from a storage location which is computed from the magnification and centering settings. 4.3 RESULTS Figures 6 through 13 show how the on-line display looks. Figures 6, 7, and 8 show an unmodified contour pattern for a grid array computed by taking the radial distance from the center of the grid. Three different magnifications are shown. Figure 9 is a double exposure of the same set of contours with a portion erased. Figure 10 shows the contours from 9 with a trough added (drawn in with the light-pen). Figures 11 and 12 are the same circular contour pattern, but with more contours used. Figure 12 shows an extreme example of the amount of magnification possible. Notice the relative sizes of the center square in the two versions. was drawn free hand. Figure 13 figure 6, contours for a radial field, at minimum size figure 7, contours for a radial field, at standard size (display size equal to size of tube face) 39 figure 8, contours for a radial field, at high magnification figure 9, double exposure (minimum and standard size) of contours from figure 7 with a portion erased 40 figure 10, double exposure (high and low magnification) of contours from figure 9 with trough drawn in figure 11, contours for a radial fiJeld, standard sime more contours than previous figures with figure 12, double exposure (high and low magnification) of contours from figure 11 with small square added figure 13, free hand drawing 42 Some difficulty is encountered in representing a curve as a set of points on the predetermined grid of display locations. The process of moving from the actual pen track to the grid often gives an uneven appearance to the line. This difficulty is mitigated by the fact that eras'ing ii-extremely easy and accurate, so that rough contours can be smoothed. Such work is especialiy easy at high'magnifications. In operational curve-drawing programs, considerable work could be done in constraining-lines to be smooth curves. However, even with this'relatively unsophisticated program it was possible to modify displayed patterns with ease. The final phase of the problem was to have been the use of the new contour pattern to modify the original grid values. Considerable difficulty was encountered, and this was not completed. It was felt that although this two dimensional fitting of a surface to the contour pattern is difficult, it is solvable, and contributes little to a demonstration of the usability of isopleth modification for man-machine interaction in meteorology. In an operational system using the approach described, it would be mandatory to have a self-maintaining, highcapacity display, such as that on the TX-2. Further import- ant additions to the isopleth display and modification would be the overlaying of other maps which could not be erased or modified. Such overlays could include geographic maps to indicate location of the pattern, station models (as normally 43 plotted on weather analyses)-for the purpose of checking of analysis validity, or, in the case of forecasts, a continuity map (a map of the contours of the initial value of the quantity being forecast). Further developments must also be attempted in the use of topological and mathematical constraints on contour patterns. Sutherland's work in automated design achieved great flexibility through the uise of such constraints. The application of similar principles to contour patterns offers great possibilities, but will require considerable labor. 44 5* THE WEATHER CENTRAL OF THE FUTURE Using the approaches which have been described thus far, what should a weather central of the future look like? First we will look at the problem from the standpoint of equipment. 5.1 EQUIPMENT One of the prime concerns in our system must be that of dependability. This requires the use of redundant compon-- ents so that if component A fails, component B may take on its job, slowing the system to half speed, but not stopping It. In the choice of central computers, we have an attrac. tive set of possibilities. Present computerized weather use the 7090, so that much programming is available for this machine. It is most desirable to retain the ability to use presently available proGrams, since a massive amount of labor was involved in their development. However, the use of two 7090 or 7094 computers is an unreasonably expensive proposition, unless such a computation facility were to be used by many agencies. Our set of choices has been increased by the recent introduction of the 7040 and 7044 computers which are considerably slower and less expensive than the 7090 or 7094. The real advantaGe of these new computers is that certain configurations which are available make them nearly identical to a 7090, except for speed. Comparative cycle times are; eight microseconds for the 7040, 2.5 microseconds for the 7044, 2.2 microseconds for the 7090, and 2.0 microseconds with instruction overlap on the 7094. 45 Thus, through a building block approach, almost any gradation of computing power and expense can be achieved, with-m out making the program library obsolete. This allows us to use redundancy without making a phenomenally expensive system. The most likely choice appears to be the use of two 7044's, each with 32,768 memory locations and extended instruction set. Careful study must be made of the cost, access time, reliability, and volume of storage necessary. At present the use of IBM 1301 disk files along with magnetic tape storage for lower speed capability seems to provide the best approaoh, but in the future new types of memories (such as thin films) may make new approaches more attractive. Redunw danoy considerations require the availability of at least two disk units, accessible by all the central computers, and possibly by direct data devices. Besides the necessity. for large volume storage due to the quantities of data to be handled, such storage .also becomes necessary in the efficient use of time sharing and the storage of programs which must be easily available. Another problem occurs because self- maintaining displays often require considerable storage, sometimes associated directly with the display'equipment. An example of this use of storaGe is the SM-2A cathode ray tube display system developed by Laboratory for Electronics. This system uses a flexible disk to store the elements to be displayed on the tube face. 46 As more economical means are found for storage of eleo.. tronic information, it will become practical to stop turning out paper products. The reasons for taking such a step are found in the increased flexibility of operations which can be performed on electronically displayed information, and in the ease of retrieval of such information. For example, it may be possible to have all the grid maps for the past year stored on magnetic tape, and to display on a console a set of contours of the change between two grids by simply asking the machine to display one minus the other. Such an idea may take a while to automate, but it results in a man being able to use the cathode ray tube and light-pen as pencil and paper, with the direct assistance of the computer in retriev-irn information and doing mathematical and meteorological operations. The equipment must also include at least two interac-' tion consoles, of the type described earlier. The cathode ray tube must have the ability to display large amounts of information rapidly enough to avoid flicker, for many meteorological displays will contain considerable detail. Equipment will also be needed for the direct input of tele-type data. The use of a satellite computer or special device for this purpose can invalidate our reliability arguments if it is not redundant. Thus, it seems most prac- tical to time share the input processing on the main computers, rather than squander money on redundant inputprocessing computers. At any rate, the initial data collection device must be extremely reliable, for its failure can halt the whole system, 5*2 PROGRAMMING Ultimately it is hoped that weather centrals will become entirely computer aided. All the operations should be able to be carried out by the computer and several meteorologists workinG at consoles. No paper of any kind will be necessary. All necessary information will be stored and retrievable by the computer. Such information will be easily displayable on any of the consoles, in any magnification desired, and in relation (overlayed, subtracted, added, etc.) to other information. During operational forecasting runs one console will be operated by a data expert. teletype data as it Before him will be displayed is normally printed. However, instead of being required to examine every report, the data expert will see only those reports or bulletins which the machine finde difficult to handle. This man will also monitor, with tho machine's help, a display of the available coverage. If any area is notably deficient, he wil recover bulletins from that area and see why no data was available. If garbling was the problem, he will correct sufficiently to allow the machine to recover the data. If no reports were available from that area, he will request data from his source stations. Another console will be operated by a meteorologist who will monitor the physical consistency of the data. He A 48 will look at vertical plots of reports which have seemed unreasonable to the machine. He will monitor the analysis process, and examine reports which the machine feels to be in error. When such reports also seem in error to him, he will request that the data expert examine the source report for garbling. Finding none, he will look at the vertical consistency of that report. Finally he will have the choice of accepting or discarding the report. The meteorologist will also aid the machine in the recognition and traoking of phenomena, such as hurricanes, fronts, the jet stream, highs, lows, etc. At forecast time he will monitor the actions of the foreoasting scheme, making sure that unreasonable patterns do not result, and aiding in the placement of the phenomena which he helped to recognize during the analysis process. Once he is satisfied with the forecast, he will give his approval, and the data will be stored until transmission time when it will be retrieved and transmitted automatically. This description has been intended only to indicate the kinds of thingfs that we can hope to achieve. Actual use of such a system will probably not be limited to operational problems, and may not be exactly of the sort described. Admittedly, the system described is the work of many years. CominG back to the present, it is necessary to examine a step by step process whereby symbiosis may be instituted in weather data processinG, without completely discarding 49 present systems. As time progresses, the success or fail- ure of the initial steps will indicate the degree to which a more advanced symbiotic system is useful and realizable. The first step must be a large one, but not as larGe as might seem to be indicated by some of the earlier dis-. oussion. Redundancy of components is not a necessity. It has been presented here because it seems foolhardy to build a powerful system upon which one may become dependent, if it is not also to be reliable. connected with symbiosis. However, redundancy is not Symbiosis may be achieved on present weather central computer systems (7090 with disk file and maGnetic tapes) by the addition of a man-computer communication console and some programming. To date, the prooessing area which seems to offer the Greatest possibility for improvement by human intervention is the initial processing of data, i. before forecasting begins. Thus, a good initial step might be to add human intervention to the programs in this phase, and develop and test a time sharing system for them. , As the time sharing programminG becomes more general, symbiosis may be extended to all aspects of operational data prooessin. If real-time data input is developed, this too may be time shared. At this point it will be feasible to begin to automate jobs which have not been put on the computer because a considerable amount of human labor must be utilized. Such jobs can be computer aided throuGh the use of the interaction consoles. As this development of the use of symbiosis 50 in all phases of forecast1.ng nears completion, there will be no need for paper products (maps, teletype printout, etc.). To develop the use of symbiosis to this degree will require the use of sophisticated information storage and retrieval techniques, for we will wish to have available both data and programs in great quantity and variety. It will also require the development of a language for speoifyingr the data fields and operations with which we wish to achieve some goal. Thus, although the ultimate system may be presently unattain-. able, the first steps can be taken easily and still provide necessary increases in the quality of operational products. REFERENCES Computer Charaoteristics Quarterly, Adams Associates, December 1962. Coons, S.A., "An Outline of the Requirements for a Computer Aided Design System," Collected Papers of the Session on Computer Aided Design, Spring Joint Computer Con-ference, Cobo Hall, Detroit, May 23, 1963, 14 pages. Fawcett, Edwin B., "Six Years of Operational Numerical Weather Prediction," Journal of Applied Meteorology, vol. 1, no, 3, Sept. 1962, pp. 318 - 332. Grisoff, S.F. "The M.I.T. Time Sharing System, A Preliminary Report, IBM memorandum, April 16, 1962. Licklider, J.C.R., "Man-Computer Symbiosis," I.R.E. Transaotions on Human Factors in Electronics, vol. HFE, March 1960, pp. 4 - 10. Licklider, J.C.R., and Clark, W., "On-Line Man-Computer Communication," Proc. of the Spring Joint Computer Conference, San Francisco, Cal., vol. 21, May 1 - 3, 1962, pp. 113 Loewe, - 128. R.T., Sisson, R.L. Generated Displays, pp. 185 - and Horowitz, P., "Computer Proo. of the I.R.E., Jan. 1961, 195. Reference Manual, IBM 7040 and 7044 Data Processing Systems, I.B.M. Corporation, 1961. Sutherland, I*E., "Sketohpad, A Man-M4achine Graphical Communication System," PhD Thesis, M.I.T. Dept. of Electrical Engineering, January 1963.