--- title: "Biological Software" authors: - "Erik Sandewall" doi: "" url: "https://ijcai.org/proceedings/1979/164" pdf: "https://ijcai.org/proceedings/1979/0164.pdf" venue: "Proceedings of the Sixth International Joint Conference on Artificial Intelligence" year: 1979 pages: "" track: "" keywords: --- # Biological Software Erik Sandewall *Proceedings of the Sixth International Joint Conference on Artificial Intelligence*. ## Abstract The robot in the common science fk:tion llnoftl appean u a man on the surface, but is : built mechanically from wheels ·and levers ·inside. In A.I., we usually visualize an A.I. system as having a similan structure: it communicates in man's language (English), .or performs other taska which make It appear man-like, but it is in fact a large program, written in a programming language, and executed under ·· a common time-sharing system. ## Full text ## BIOLOGICAL BOBTWABB Erik. Sandewall: Informatics Laboratory LinkiSping University Linkoping, Sweden Abstract end lntrodtutltm.~ The robot in the common science fk:tion llnoftl appean u a man on the surface, but is : built mechanically from wheels ·and levers ·inside. In A.I., we usually visualize an A.I. system as having a similan structure: it communicates in man's language (English), .or performs other taska which make It appear man-like, but it is in fact a large program, written in a programming language, and executed under··a common time-sharing system. The present paper argues on the contrary that A·.l. systems ·S!t.ould not need to appear man-like, and that it is necessary, both from the A.l. point of view and from; the software engineering point of view, that the next level down in the1·A.I. system has quasi-biological properties, such as the ability to reproduce. R1marll: du.1 to spau ltmttcttons, tAts pap~r Aas b11n fJJTtttln In tJ cond1ns1d st,ll. 1. R1productton ts a m1cllantsm for a sp~et1s to :t.ul4pt. It is natural to make analogies between the biological system that forms the substrate of natural intelligence, and the computer hardware/software syatem that ~forma the substrate of A.l. Among characteristic properties of organisms we find homeostasll and the ability to reproduce. Do they have counterparts in software? The common game of writing programs that · aeate many identical copies of themselvei, misseJ an important reason for biological reproduction: it is a m1cAtJntsm for adaptatton, from the point of View of the species. Human reproduction involves not· only copulation and. child-birth, but also · two decades of work for training the new. individual (sometimes called ·social reproduction· by sociologists); In the training process, the new person inherits some knowledge, but also selects, reviews and reorganizes his or her intellectual heritage. Imagine how much progress· there would be if humans lived forever and never reproduced! 2. Softwau also n11ds to adapt, tn 11spons1 to claantlng usiT · n11ds. Conventional software does not adapt CJutonomou.sl' to an' stgntjtcant l:ttlnt:.:· U Is e~daptld by a programme!! who changes the program (·malnt1nanc1·) or throws it away and writes ·a new one. Future software systems with a more ·complex behavior will have much bigger needs to adapt, and will be harder to change from the outside. They should therefore be enabled to change without programmer intervention, in response to requests from the end user. They can do so either by modifying their internal state, or by creating a modified copy of themselves. Botlt of tAISI mltltods are vtabl1 tf tAl progTa'M1fllng t1clantqu1 of conceptual programming <ts us1d. Remember that: adaptation was a topic of interest in earlier A.l. work, but it has gone out of fashion because it seemed one could only do trivial thinga with it, such as adjusting numerical parameters. What one would really like to do was to adapt programs, but that seemed too hard, and had to wait for automatic programming to happen first. But there is a trend In several areas of software to use specialized application languages (implemented either as interpreters i.e. general, highly parametrized programs, or as program generators) to develop applications. Examples of this trend are: - commercial program generators for reports, screen layouts, data. ·base queries, etc. - the system for information-flow applications, developed in our group (ref. 1). - special-purpose languages in A.I. research e.g. grammar languages. These are signs of an emerging software technology where computer applications are implemented by selecting a small number of general-purpose r tools, combining them in appropriate ways, and providing them with a description of the applicration in (several) specialized application languages. This style of programming has been called cone~pt'U.al programming In ref. 2,3,i. Self -modification in a system should clearly be relatively easy .tf it can be performed by modification of parameters or application-language expressions, as compared to direct self-modification of a conventional program. This is both because easier. "maintenance" is one of the reasons for specialized languages, and because end-user requests for modification. are naturally expressed in application-language term" a. Should an A.!. S"jSttm 'livt' fortvtr1 In other words, is internal self-modification sufficient as an adaptation mechanism, and is reproduction unnecessary? The answer is that perhaps it is not practical for software to adapt forever. · For adaptation, new knowledge has to be added to a system, which means that other knowledge has to give way, otherWise . the system will grow· indefinitely. But whenever knowledge is removed, one encounters a well-known but yet un-named problem which I propose to call thl dtlttt probltm, for example in the following simple form: suppose A has been asserted in a data base, and B has also been asserted by forward Inference using "A implies B", and later A Is to be deleted. Should B be deleted as well? There might be other, independent support for B, so that it should not be deleted. The problem is well-known on many software levels, from conventional garbage collectors to semantic networks. The conclusion Is that for a system to be indefinitely adaptible, it must contain' a considerable overhead of information about the details of its own insidet, such as back pointers from B to all its independent supports, In the example. Reproduction offers an alternative, with copyingr garbage collection as an example on the low software level, and training of off -spring as examples among animals and (I conjecture) for A.l. systems. 4. Rtproduction talus man1 forms; ours is a biological txception. Reproduction in mammals has a simple structure: two Individuals togeth1:r generate an offspring which grows up continously. But butterflies, ants, fish, jellyfish, frogs, and funghi offer a rich repertoire of more complicated schemes, where the organism exists in more than one physical form during the reproductive cycle. In several projects In our· research group, I have seen needs for similar, non-trivial schemes for reproduction in software. - Bacl!.ground: som1 of tla1 group's worlt assum1s tltat th1 programming ttchniqu·s that havt b11m dtvtlop·d in A./. r1starr:la, art a significant spinoff r1sult of that rtstarcla, and should b1 used for ot!a1r pu.rpos1s as wtll. Two specific cases are of interest: - .A) Ust of tht lnttrlisp S"jSttm for dtvtlopmtnt of pilot vtr stons of application programs .. (A pilot veraion is one which can easily be changed until the end users are happy). Development of the initial program went smoothly, but the subsequent user-Initiated modifications were clumsy to make" Analysis: . when conventional languages are used, one does debugging in the production environment (i.e. using the regular compiler). The Interlisp system provides a "dry-dock." environment for program development, which implies that a non-trivial effort is required to take the program into and out of the "dry dock.". This is worthwhile if a lot of work has to be done there, and in particular if the program never really leaves the development environment, as is often the case in A.l. research. But the same programmer support did not seem worthwhile as the programmer's effort for transfer from development environment to production environment was much bigger than the effort of the update Itself. The problem was solved by allowing·: each application to exist In two forms: a development system and a production system. The development- · system was the one that really 'adapted', and it had the ability to create corresponding production versions. This solution embodies the following basic attitude: a programming system should not be like a street-front clinic, where a program walks in to be operated on (for example, for being compiled, or debugged during - a debugging session), and then leaves again. It should instead be a permanent dwelling for the program. The programming system plus the program it accomodates, should be viewed as an organism, which contains a fair amount of knowledge about itself, and In particular has the ability to generate systems which are similar to itself although tuned for production use - like a queen bee does. - B) Modelling informatton-jlow s'Ystems. Many data processing applications can be characterized as Information-flow systems: there are specialized work-stations for different people or roles in the organization; Information packets ("transactions") are generated in such stations, sent in channels from one station to the next, accumulated in files that are local to each station, and the task of the person at a station is to review the information that flows by, take actions, add more information to the transactions, and pass them on. Recent work in our group has resulted in a prototype software tool which allows the various aspects of the Information-flow application to be described In specialized languages, and which allows one to do pilot execution of the system, work on a generator of production systems is in progress. The system is designed with the programming techniques that are usual In A.l. programing, such as strongly parametrized programs, handles, and data-driven procedure calls. In the terms of the previous sections, this system is "organism-like" in that it can create modified copies of itself, namely copies for each of the work-stations. These specialized production-phase copies have been obtained by selecting subsets of the parameter structure, and by specializing programs through "smart" compiler macros. These and other applications in our group are using a general-purpose tool for accomplishing system reproduction, called ACTEMAN (ref. 5). 5. Knowledge acquisition. The reproduction in these two examples as presently implemented, does not involve any selective acquisition of knowledge by the off-spring. (There is however a rea' need already to have that). The major reason for the examples Ia to illustrate that multiple 'life-forms' and unusual (for us) reproduction structures may be useful for software systems. With a system which is organized along the lines of conceptual programming.. as mentioned above, knowledge acquisition by the young off-spring could be organized as follows: - let the ancestor generate a basic system (for exal':'lple a fresh instance of the Lisp system), and load the appropriate general programs into it. Also let the ancestor transfer the appropriate parameters and datadriven procedures to the descendant, but usually only a subset · of what the ancestor "'knows" - let the descendant engage In work., usually by serving a human user, where in doing so one identifies missing parts In its structure and obtains the information by communication with senior computer systems, (in particular, systems similar to the "deveopment phase" system described above, which contain extra knowledge and which specialize in training new systems), and only in hard cases, communication with a human programmer. In such an architecture, it Is possible but not natural to let the computer systems communicate between themselves In human natural language It is much more natural to exploit the ,properties of the computer medium, and let them communicate in terms of program segments, data structures, and so forth . (This does not preclude that some of the characteristics of naturallanguage communication will stitt :>e needed) 8. We can now proceed to the other issue mentioned In the Introduction to this paper: T !at emphasts in current A./. resuzrcla on s'1st1ms wit/a laumanotd belaavtor. Its main historical reasons are: - A) The training issue: It has been argued that intelligence presumes knowledge, and that the computer can only gain knowledge by talking to people and/or being among people, as a robot. - B) The usefulness issue: it has been argued that an intelligent computer system can only be useful to us people, if it can talk. to us in our language (which is assumed to be natural language). The training argument addresses the wrong bottleneck. It is true that a lot of work will have to be spent by people for transferring knowledge to computer systems, and it is conceivable that it will be facilitated If natural language may be used. But already scores of programmers have as their profession to transfer knowledge and ability to computers. The significant problem is instead that computer systems can not transfer knowledge to each other. If one computer system could train others, ~hen it would not matter if the initial knowledge transfer to one computer 'system had to be done In a formal language The reproduction chain can start when we have built systems which have the necessary properties for being useful, and for generating useful offspring, recursively. In order to start it, we ourselves wilt have to understand the reproduction process fairly much In detail. It then seems more convenient to create the first generation by performing manually the operations that will later be done by the ancestor, rather than create a natural-language understanding capability only for the purpose of the first system generation. The usefulness argument (B) is valid if we consider current research results as given, and search for possible applications. But suppose Instead that we would focus on currtnt data processing (D.P.) applications, and ask what role programmed intelligence could play in them. The role might not be dominant: tht resulting thought product is not necessarily a computer intelligence incarnated Into an application. A car with a micro computer in it is well described as a car and not well described as a computer, and intelligence might be very useful in a similar, subordinate role in the D.P. system. 7. There is a significant area of common interest for A .I. and software engineering. In A .I., conventional programming has traditionally been considered - as an intelligence-requiring activity that is observed to occur in real life, and which therefore Is a candidate A.I. application, and - as a part of the support work for an A.I. lab and a temptation for the graduate student to waste his 4. time on It Is then assumed that software engineering and A.I. are disjoint areas. The arguments of the present 1paper imply on the contrary that - systems which adapt, internally and by reproduction, are necessary both for A.I. and for future software engineering - conceptual programming techniques may facilitate design of large and complex systems (which is a topic of common interest. for A.I. and S.E.) as well as adaptation and reproduction. - such systems, performing conventional D.P. tasks, are a likely early application area for A.I. research. But intelligence can not then be obtained as an add-on: the system has to be built from the start using some of the programming techniques which are common in A .I. 8. How does this fit in with other work in A./. Hewitt's actor concept is clearly related in the abstract sense that he also talks of structures formed by individuals with a local autonomy and initiative. However, he has mostly worked with actors of very simple structure, In particular as an elementary object in a theory of computation. We emphasize systems which are complex enough to contain intelligence; systems which are pseudo-individuals. Marr's work. in computer vision has already started a trend to take the biological foundations of intelligent systems seriously again ## Acknowledgements Jim Goodwin and Sture Hagglund have contributed strongly to the ideas expressed in this paper. ## References 1. I. Erik Sandewall: A description language and pilot-execution executive for information-transport s'Ystems. Proc. Very Large Data Bases )conference, 1979. 2. Terry Winograd: Five lecturu on artiftclal intelligence. Stanford A.l. Memo, 1974 3. Erik Sandewall: Some observations on conc-fJtual programming. Machine Intelligence 8, 1977 4. Bob Wielinga: A I Programming Methodolon. Proceedings of the AISB/GI Conference on Artificial Intelligence, Hamburg, 1978 5. Erik Sandewall: Self-organizing lnformatton and Operations for Reproduction tn Distributed Programming S'Ystems. Internal report, Informatics Laboratory, Linkoping University, 1979.