File Information
File: 05-lr/acl_arc_1_sum/cleansed_text/xml_by_section/metho/75/t75-2021_metho.xml
Size: 31,317 bytes
Last Modified: 2025-10-06 14:11:11
<?xml version="1.0" standalone="yes"?> <Paper uid="T75-2021"> <Title>Hewltt, Carl et..al. &quot;Actor Induction and Meta-evaluation&quot; Conference Record of ACM Symposium on Principles of Programming Languages. Boston. Oct, 1975. &quot;</Title> <Section position="2" start_page="0" end_page="0" type="metho"> <SectionTitle> Carl Hewitt Abstrapt </SectionTitle> <Paragraph position="0"> This paper is a spin-off of our work on actors. * We have worked out a dictionary for translating between what M.insky 'et. ah are saying about frames and what we are saying about actors. Using PT.ASIV\[A \[PLANNER-like System iV\[odeled on Actors\].we can demonstrate important rela.tionships between the Minsky-frames and the PLANNER-like form~.lisms. PLASMA does not .use the Q A-4 context mechanism (Rulifson et. al. 19'12): instead It uses explicit tags in assertions to keep track of the state of affairs m various situations. One problem with the.QA-'i context mechanism is that the problem solver is forced to attempt to propagate all changes in the situation immediately on a frame, shi(t since otherwise inconsistent information will be inherited from the previoussituation.</Paragraph> <Paragraph position="1"> Another problem with Q A-4 context meclhanism is that it is sometimes difficult to reason explicitly about various ~ituations using it because situations \[frames\] are not explicitly part of the assertions and goals. Events that are Viewed from several:different viewpoints \[as in a murder mystery\] are diffic01t to handle. Also It Is difficult to retrieve the appropriate prior sltuat!ons from memory to aid In recognition tasks using Q A-4 contexts. However, without the example of the Q A-,i contexts to guide us, we could never have realized holy to deal with these, problems using' tagged assertions.</Paragraph> <Paragraph position="2"> The context mechanism in' CONNIVER was modeled on the one i~n o_..A-4.</Paragraph> <Paragraph position="3"> Actors make a contribution to the &quot;dec\[aratlve-procedure&quot; controversy in that they subsume both the behavior of pure procedures \[functions\] and pore declaratives \[data structures\] as special cases. In this paper we use actors to investigate the question of how to do procedural attachment to frames \[McCarthy 1969, * Minsky et. al. 1974\]. Actors. provide an approach to solving to the problem of how to attach procedures to frames in such a way thai the appropriate procedure is invoked with the inherited knowledge of the frame within Which the procedure is incorporated, as well as newly introduced knowledge from other frames. Our approach also incorporates the insight gained from PLANNER-like formalisms \[Hewitt 1969, 1971; Rulifson 1972; .Davies: 197'2; Sussman and McDermott 1972, Hewitt et. al. 19&quot;/3\] for the'procedural embedding of knowledge. We introduce stereotypes as an actor version of a frame theory. A stereotype consists of the following parts: a collection of eharaoterist;io objects characteristic x:elations for those objects plans invoked bY world directed invocation for transforming the objects and relations qa Overview /\]pologyt. The schemes proposed herein are incomplete in many respects. First, I often propose representations without speci\[yi.g the. processes that will use them. Sometimes I only describe properties the structures should exhibit. \[ talk about markers and assignments as though it were obvious how they are attached and linked; it is not, &quot; Marvin Minsky i~74 ~ FRAMES like ACTORS are difficult to define and motivate.</Paragraph> <Paragraph position="4"> Through the dint of much effort \[C,reif and Hewitt 1975\] actors are now becoming quite well defined \[even axiomatized!\].. In this paper we make a stab at similarly starting to pin down and unify various notions of &quot;frames&quot;. We have observed a tendency for people to get hung up on extraneous details as they attempt to understand these difficult concepts. The*notion s of default values and expectations in particUlar are a cause of diff.iculty. For example expectation * is often confused with scientific' prediction. In attempting to sort all this out, we have found the methodology of CONJECTURES and REFUTATIONS as expounded by Karl Popper tobe of great value.</Paragraph> <Paragraph position="5"> The relationships between the assertional frames of McCarthy and the object frames of Minsky are explored m this paper using PLASMA. PLASMA allows us to demonstrate a kind or isomorphism between assertional frames and object frames. Using PLASMA accomplishes procedural attachment for both kinds of frames and shows how to move between object and as~ertional frames.</Paragraph> <Paragraph position="6"> Assertional Frames anc~ Semantic Nets.</Paragraph> <Paragraph position="7"> The example given *below was originally used by Terry Winograd to illustrate the properties o( Semantic Nets in his.</Paragraph> <Paragraph position="8"> informal survey entitled &quot;Fiv.e Lectures on Artificial Intelligence'. Scott Fahlman is currently considering how to build a special purpose machine to solve such problems. We first give a formulation in terms of assertional frames. The addition of assertions of the form (... never-are ...) was inspired by Grossman \[1975\]. It is interesting to speculate whether the formulation in terms of assertiona\[ frames is as plausible a psychological model for humans as the formulation in terms of semantic nets.</Paragraph> <Paragraph position="9"> Next we formulate the example in semantic networks.</Paragraph> <Paragraph position="11"> * Relationship of Worlds. to Contexts We conceive of worlds {Hewitt el. el. 1973} as actors that organize a body -of knowledge for efficient use. Worlds are generalizations of the tree-structured contexts developed for .QA-4 (Rulifson el. el. 19/I) as a generalization of the data bates of PLANNER-69. * .</Paragraph> <Paragraph position="12"> &quot;/J Frame is a coitection o\[ questions to be asked about a hypothetical situation; it apeci\[ies issue, to be ralsedand methods to be used in dealing with them.&quot; &quot;In \[oct we shall&quot; con~ider the idea that the .frame terminal, \[}.or a scenario str~ctureJ ore exactly those questions \[commonly associated wilh it\[.&quot;' Marvin Mlnsky 1974 Worlds have many of the behaviors attibuted to 0bject frames {Minsky 19&quot;H} and assertional frames {McCarthy and Hayes 1969}. In particular worlds exhibit a very flexible form of inheritance of attributes which is realized through the mechanism of message passing.</Paragraph> <Paragraph position="13"> Progressive Refinement of Plans Part of Minsky's birthday party frame includes a PLANNER-style general plan for getting ready for a birthday pal&quot;ty. It illustrates how PLANNER-style plans.are a natural component of frame systems. In PLASMA the general plan can be expressed as follows: We use the above general plan as a template to write more particular plans tailored to special circumstance by progressive refinement of plans \[Hewitt IJCAI-?I\] of'the plan. This technique is an incremental glorified version of iqline substitution of procedure bodies for their invocations which has been ge.neralized to deal with pattern directed invocation. Cheatham and .Wegbreit \[19&quot;/3\] and Burstall and Darlington \[1975\] have studied the problems involved in the absence of pattern directed invocation.</Paragraph> <Paragraph position="14"> We will suppose that early one morning that Marvin has been told that there will be a birthday party for Seymour the next evening. The above plan becomes specialized as follows: As Marvin rehearses* his planned' actions for the day his plan * becomes further specialized. Eva, who is a friend of Seymour, probably will have suggestions as to what might be a suitable present for Seymour. Also, a party for Seymour Is not the kind of affair that guests dress up for. These considerations result in the following refinement of the plan: ..</Paragraph> <Paragraph position="15"> The above plans were formulated between yawns as Marvin awakened the morning, of the party Procedural Attachment in Assertional Frames &quot;/! situation s is the complete state of the universe at an instant of time. We denote by Sit the set of all situations. Since the universe is too large for complete description, we shall never completely describe a situation; we shall only give facts about situations. These facts will be used to deduce further facts about that situation, about future situations and about situations that persons can bring about from that situation, * ..o ~Ve shall further assume that the lows of motion determine from a situation all future &quot;situations. {This assumption is difficult to square with quantum mechanics, and relat~ely tells us that any assignment of simultaneity tO events in different places is arbitrary. However, we are proceeding on tEe'basis teal modern pkysics is irrelevant to common tents deciding WEar to do, and in particular is irrelevant to solving tee 'free .will problem':}&quot; John McCarthy 1969 In addition to making the assumptions listed above McCarthy and Hayes also assume the existence of a functional which maps a global state in .to the &quot;nexC global state. Although.we shall use tags on assertions and goals to relativize them to particular situations, hypotheses, and contexts we do not want to make any of the global assumptions of McCarthy and Hayes. All of our situations will be local and completely consistent with relativity and. quantum mechanics. For a discussion of the issues involved in this decisiott for the actor model of computation see Greif and Hewitt \[19'/5\]. The Pattern Directed Invocation \[Hewltt 1969\] incorporated in PLANNER-like systems accomplishes proced~iral attachment for assertional frames \[McCarthy 1969\]. 'This procedural attachment is carried down to the level of the quantificatlonal calculus *where we have shown \[Hewitt 1975a\] how' the logical operators of the quantificational calculus {V, 3, implies, ^, v, -, etc.} can be behaviorally defined as actors. We use the following kinds of tags in assertions to relatlvlze the assertions in the desired manner:. : We introduce stereoty'pes as an actor version of a frame theory. Our notion' of a stereotype incorporates ideas from assertional frames \[McCarthy 1969\], act~ worlds \[He'witt 1969, 1973\], world-directed invocation \[Hewitt 1969; 1971, 19&quot;/3; Stansfield 1975\]; social frames \[Goffman 197st\], and object frames \[Minsky et. al. 19'/'t\]. A stereotype consists of a set of the following parts: a collection of characteristic objects characteristic relations for those objects a set of plans invoked by world directed invocation for transforming the objects and relations %* It seems that many of the behaviors attributed to frames by Minsky can be realized by stereotypes. The characteristic objects of a stereotype correspond closely to the slots of a. Minsky frame and the characteristic relations of a stereotype correspond to the constraints of a Minsky frame. Minsky calls simp\]e ,ma.ry characteristic relations markers. We instantiate stereotypes somewhat differently from the way in which Minsk? instantiates object frames. Our approach is to plug in definite candidates for all the characteristic objects of a stereotype that we can and use anonymous objects for the rest. Defaults are done as assertions tagged tO indicate that they are defaulted so that they can be easily displaced if an anomaly develops. Stereotypes communicate by making assertions in the data base and by world directed invocation which is a generalization of pattern directed *invocation in which the invocation is done on the basis of a fragment Of a micro-world instead of a single assertion. Inheritance of attributes is done using the message passing of actor semantics. If a questions cannot be answered directly then parts of the job.are &quot; delegated.</Paragraph> <Paragraph position="16"> Prooedur~'d Attachment for Stereotypes * ~ We have found it useful to incorporate ~ types of worlds \[called UTO\]PIA and REAI&quot;.ITY\] in .every problem solving situation in order to incorporate goal &quot; orientation. The UTOPIA-REALITY machinery also enables us to incorporate PLANNING into problem solving using &quot;islands&quot; \[Minsky 1963\] as stepping Stones. The utopia of one problem solving situation is taken as the reality of another.</Paragraph> <Paragraph position="17"> &quot;It will be worth a r~latively enormous effort.to find suck 'islands' in tee solution of eomplez problems. Note teat even i/ one encountered, say, 106 failures o/ such procedures before success, one would still, have gained a factor of perhaps I0 tO in over-all trial reduction! T/uLS practically any ability at all to 'plan/ or 'anaiy~e,' a problem will be pro/liable r if the problem is dif\[ieul|.&quot; Simple retrieval can be done using fragments of micro-worlds that consist of single patterns \[pattern-directed retrleval\]{Hewltt 1969}. For example (find (=apt is-an-apartment-i*n Cambridge) ;find an al~rtment in Cambrid~re (then&quot; ;then (refute (apt is-nn-acceptoble-apartmenr) ;find something wren I with tee proposed apartmen! (else: ;else (move-irdo apt)i)) ;move into tee apartment (else: ;~Ite (move-to Arlington)}} ;move to flrlinKton The use of &quot;then:&quot; and &quot;else:&quot; continuations seemsto solve the scoping control problem which had been plaguing PLANNER-like languages for some time. CONNIVER attempted to solving the scoping control ' problem by introducing possibility lists and Landln-style non-hierarchical gotos. However possibility lists proved to *have several deficiencies* They introduced side-effects into the. basic communication mechanisms in CONNIVER which, made it difficult for users to debug their programs since doing a try-next operation to</Paragraph> <Paragraph position="19"> print the next possibility destructively interferes with the operation of the programs being debugged. The other, basic communicatioll mechanisms of CONNIVER similarly have intrinsic'~ide-effects built into their very structure.</Paragraph> <Paragraph position="20"> &quot;Their \[Sussman and McDermott\] sohttion, to give the user access to the implementation primitives of PI,/INNER, is however, something of a retrograde.step (what are CONNIVER's semantics?), although pragmatically useful and important in the short term, t} better solution is to give the user access to a meaningful set of primitive control abilities in a~t explicit * representational scheme concerned with deductive cqntroi.&quot; Pat llayes 1'974 Nested continuation control structure gives us the ability to influence or control any decision to the extent we desire. For example, in case of the apartment finder above, we can explicitly communicate complaints as to why a particular apartment is unacceptable in order to try to infiuence the selection of further proposed apartments. Use of nested continuation control structure enables us to have the ability to control all o.f the decisio.ns made while still retaining the high level goat oriented, nature:of PLANNER. PLASMA is able to accomplish this by basing its semantics on actor message passing and slightly changing the syntax ofPLANNER-?I, The change in syntax provides us with natural places to incorporate the (:ontrol information and enables us to avoid the gratuitous side effects !n PLANNER-?L Unification of Pattern-Directed Invocation We do not want to have .to explicitly *store every piece of knowledge .which we have but would like to be able to derive conclusions from what is already* known using procedures. Using the distinguished symbol when with the syntax (when trigger consider-trying b.ody) or completely equivalently using the distinguished symbol to with the syntax (to trigger consider-trylng body) creates a plan \[high level goal-oriented procedure\] that can be invoked by pattern directed invocation by a trigger Which matches trigger. The following are all special case plans which are defined in terms of the above general pattern directed invocation machinery. The process of. invoking plans in worlds is controlled by recommendations made to the world when the plan is put in the world and by recommendations made at the site where a request is made of the world to achieve a particular goal. We envisage that problem solving would begin in a world with an initial class of plans. In many cases most of the plans that are used in. the ultimate solution of the problem need to be constructed by other plans during the problem solving process. This is ilTustrated in a limited way by the domain of logic which is given as an example in Hew(st \[IJCA\]-75\]. Pattern Directed Invocation Using AnonymousObjects One important way in which plans can communicate is through making assertions, erasures, denials using worlds which they share in common* Another important means of communication is through pattern directed invocation. An important technical problem in implementing pattern directed invocation is how to solve the problem of matching the invoking pattern with the pattern in the trigger of the *pattern of the plan being invokecl..PLASMA uses anonymous objects to solve the .problem. We assume the existence of a generator capable of generating new anonymous individuals anon I, anon 2, etc. which have never before been encountered. To show the utility of such a generator consider the problem of proving (x c_. z) \[x is a subset of z\] where we have a world which Contains:</Paragraph> <Paragraph position="22"> \[demonstrate-C/-= ;define the plan'demonstrate-c_ to be (to (demonstrate (=a~ =c)) try ;to demonstrQte that a is a subset o'\[ e try (demonst#ate (a C- fib) ;to demonstrate ihata is a subset of another set \[eall'~t b/ (then: ;then (demonstrate (b C- C) * ;demonstrtite t.hat b is a subset of c (using: set-theory))) (using: set-theory)))\] &quot; The problem is solved by &quot;wishful thinking'* i.e. reasoning within a hypothetical world. Unfortunatelywe do not have the fact (x C/ z) explicitly given to us and so must do .some computation. We note that we have a plan whose trigger (=a c: --c) matches what we are trying to achieve and so turn control over to it to see what it can do. In order to find b_ such that (x C- b_) we let b_ be anon I which is a never before encountered individual which we wish to have certain properties. Then we note that anon I might be w. But we are unable to demonstrate \[w C/__ z) so we reconsider and see that anonl might be y. We successfully demonstrate* (y c_ z) and so the problem is solved.. There is not much that can be done with a line seg6~ent. However given one vertex, we of then need to be able * to get the oppos!te vertex. A plan for doing this is given below: grind-opposite-vertex (to (j'ind =v i is-opposile sv 2 alontr me) ((side s is-a-vleu~-of lace s af block i) in the-view) ((side b is-a-view-el taceb, of block i ) in the-view) ((side c is-a-t, lew-of face c of block 1) in the-view) . ((side s i~-Ieft-helaw side c) in the-view) ((side b ia-rlgEt-kelou~ side c) in the-view) ((v e is-a-Y-retreat) in the-view)</Paragraph> <Paragraph position="24"> Stereotype for Typical View of a Cube Reoognitior~ of Stereotypes ~'e make proRresl if, and only if. w~. are prepared I0 learn from our mistakes.</Paragraph> <Paragraph position="25"> Karl R. Popper Philosophically, this section builds on some work by Karl Popper on Conjectures and Refutations as a scientific methodology. More concretely it builds on a fine piece of work Ben Kuipers did on recognition of kinds of blocks \[Kuipers 1974\]. The reader is referred to KuIpers! paper for various kinds of discussion and background which will not be repeated here. The recognition problem Is to incrementally recognize the above three kinds of blocks. This kind of problem has been extensively investigated by.robot projects at the M.I.T. Artificial intelligence Laboratory, the Stanford Artificial Intelligence Laboratory, and the S.R.I. Robot Project. In order to facilitate comparison, we have fairly closely followe.d Kuiper's scenario. Our motivation in re-exploring the block recognition problem is to Investigate some techniques that have been developed by Marllyn McLennan for the very difficult. 'domain of understanding plant pictures. We wanted to investigate a technique that she is developing for resolving conflicts, between conflicting hypotheses recognition problems in another domain.</Paragraph> <Paragraph position="26"> We have modified Kulpers scenario of the recognition of a block using stereotypes by using three-dimension stereotypes Instead of the two-dimensional views which he used. These three-dimensional stereotypes are reminiscent of. some of the three-dimensional &quot;models&quot; used in the robot vision programs. We start the recognition with an initial vertex v I (see figure below), which in this case happens to be an L-vertex. Our initial hypothesis h is that the drawing represents a cube as indicated by the dotted lines:</Paragraph> <Paragraph position="28"> We have the following situation:</Paragraph> <Paragraph position="30"> We now have the following information: We assume that the visual primitives that we are using are region-oriented as well as being line.oriented so that we can directly detect and recognized the shape of certain kinds of regions.</Paragraph> <Paragraph position="32"> ((v 4 corresponds-to V a) in h) ((side I is-a-view-of face a of cube 1) i, h) The Y-vertex at the center of the figure also corresponds completely with the cube hypothesis h. A complete parallelogram face has now been observed and confirmed.</Paragraph> <Paragraph position="33"> v, &quot;'&quot; ,, .s;.,,.</Paragraph> <Paragraph position="34"> With this observation, the cube hypothesis h finally breaks i down. The anomaly occurs because the assertion that s.ide 2 is a II triangle conflicts with the characteristic relationof the cube s~ereotype that all of its faces are squares. Thus there Is .no easy to resolve the anomaly.by simply rotating the cube* However, the cube stereotype i is reluctant to give up completely so it Imagines one of its sides II shrunk down to an edge to fit the data. This reminds cube stereotype of the the wedge sterfotype, which it suggests. The triangular side, in effect, caused the recognition system to do a &quot;double-take&quot; creating i hypothes'is h' that the block isa wedge since the wed|e stereotype Is I happy with the situation.</Paragraph> <Paragraph position="36"> ((block I is-a wedge) in b') | ((side I is-a-view-of face c ofwedge 1) inh') ((side 2 is-a-dew-of fKo a of wedge 1) i~i h') At this point, with. the wedge stereotype directing the exploration, there is only one remaining edge. Unfortunately it. refutes the hypothesis h' U/&quot; I V, ec.-,~ V~ Vertex 5 ConUnued (v 5 i,-a-~rtez-of * 3) (side 3 im-a-trlangle) The observation that side 3 is a triangle conflicts with the</Paragraph> <Paragraph position="38"> characteristic relation of the wedge stereotypetheir a wedge does not have two triangular faces Which are adjacent. Again no amount of rotation will help at all. However the triangular face side 3 suggests squeezing the edge of the wedge down to a point suggesting a pyramid. The pyramid stereotype takes a look at side I and decides that the square-pyramid stereotype should be invoked, (h&quot; ohtnlnr, d-hy-refulatinn-o/&quot; h') ((block i is-tz square-pyramid) in h&quot;) ((side I is-n-i~i~w-o\[ face eo\[ square-pyramid 1) in h&quot;) ((side?. is-a-t;iew-o\[ face a af square-pyramid\[) in h&quot;) ((side 3 is-~z-f,iew-o\[ face b o J&quot; squareopyramid I ) in h&quot;). Since there is no further input data to be considered and .thus further processing yields no.refutations, hypothesis h' temporarily survives.</Paragraph> <Paragraph position="39"> T_racking the Image of a, Cube In this section we shall consider ,an example due to Mlnsky from the point of view of stereotypes. We assume.that the result of looking at the cube pictured below I Of course the situation of the cube has not changed but we have the following additional information: ((C has-color white) in situation 1) (C is-a-view-o j&quot; facef o\[ cube l) . .</Paragraph> <Paragraph position="40"> How if we imagine moving back to \[he left, we can hypothesize the view without any perceptual computation at all. We simply imagine that the new view view 3 is the same as view i.</Paragraph> <Paragraph position="41"> ((view 3 = view 1) in hypothesis l) However, moving back to the left we are surprised to find that in the new view \[view 3\] that A has changed color to while! Thus hypothesis ! is rejected. We also notice that ((floor has flecks-of-white-paint) in view3)&quot; which causes us to construct another hypothesis&quot; (situation 2 = (pain! A white in situationl)) ((view 3 is-a-IJiew-o\[ situation 2) in hYpothesis2) .</Paragraph> <Paragraph position="42"> Further careful observation and testing.does not refute hypothesis 2, so situation 2 can inherit suitably transformed attributes from situation I using the hypothesized paint transition.</Paragraph> <Paragraph position="43"> More .degn Inheritan,ce ofAttributes The &quot;frame&quot; problem of McCarthy for assertional frames corresponds closely to the problems of &quot;inheritance of attributes&quot; and view I I &quot; &quot;default values&quot; for object frames. For example the fable quoted by is the following symbolic description: Minsky about the wolf and the lamb is very close to the frame problems of McCarthy.</Paragraph> <Paragraph position="44"> ((A has-color red) in view I ) ((B' he J-color while) in situation 1) * ((E' has-color white) * in situation l) (A' |a-a-riP.w-o\[ face a o j&quot; cube 2) (B' is-a-view-of face bo\[ cube2 )~ (E' ia-a-vlew-o\[ facet o j&quot; cube 2) Now we can create situation 4 by gluing side C of cube I to side A' of cube 2 to give block I.</Paragraph> <Paragraph position="46"> (situation 4 re.~ults-frorn (glue C to A ~, in situation l)) ((block I (blocks) in situations,) ' ((ma s block I ) = ((mass cube ! ) / (mass cube2))) (((face a o,r cube i ) i.~-a-/ace.-o\[ block/) in situation4) (((face a o,r cube i ) he~-coior red) in situation4) Note that some attributes in siluation 4 such as the color of A are inherited directly from one of the cubes in situationl: ' On the other hand the mass of block I is inherited as the sum.of the masses of cube I and cube 2.</Paragraph> <Paragraph position="47"> Real World l~eoognition The principal reason why vision programs at present perform so poorly is that the amount of knowledge they can bring to bear on the seeing process is so limited. For example, Waltz's program \[Waltz 1972\] is the latest in a line of development called scene analys, which was originated by Guzman \[1968\], and pursued by Huff man \[1970\] and Clowes \[1971\]. The usefulness' of this appro;ich is called into question by the difficulty of extracting, from information about intensity and color, the near perfect line drawing that such programs require; and by the restricted nature of the line drawing represeniation itself~ Waltz has demonstrated&quot; that when a more detailed categorization of the kinds of labels is made that the number of ambiguous line drawings decreases dramatically. &quot; We greatly admire Waltz's program, .but we f~l that the apprOach that it embodies is open to several criticisms. The first is that the knowledge that it uses is in a certain sense not explicit enough. Although it contains a great deal of. information about the appearance of line drawings, this information is essentially, in a compiled form: one reflection of this ts that the structure of Waitz's program makes it inherently unable to use either explicit information about the three-dimensional form of what is being viewed, orthe many pieces of special and general knowledge that we surely bring to bear on the process of seeing.' Essentially, the only way one.can attempt to influence the program is by adding or deleting junctions from a large table of &quot;legal&quot; labellings. There is no way in which pieces of its knowledge can be pulled out and examined while tt tries to create an interpretation of, for example, a scene in which several lines are missing. Unless such knowledge, suitably embedded in a hypothetico-deductive system, can play a large part in the operation ofa vision program, we see no prospect of such a program being able to interpret the incomplete information that is the diet of daily life.</Paragraph> <Paragraph position="48"> The basic trouble with the labelling approach of scene analysis is that it is too limiting and stultifying a paradigm for vision, in much. the same way that resolution is for deduction. The fundamental * principle of resolution, that (- A) and {A v B,) together imply B, is occasionally useful, But attempting to make a uniform resolution proof procedure, to mechanize deduction in a way that cannot be very sensitive to hints, hunches, and a wide variety or higher level knowledge about the particular domain in question, is a cul-de-sac.</Paragraph> <Paragraph position="49"> Similarly, the line and vertex labels are local, predicates that are occasionally useful, and are of *some mathematical interest in their own right: but the problem of creating a uniform procedure to label arbitrary line drawing is not a central one for vision. Hence, we believe that the kind of knowledge contained in Waltz's program is probably relatively unimportant; and that the way in which it is made available there is certainly too restricting.</Paragraph> <Paragraph position="50"> The proper endeavor of vision research is to decide what k.nowledge should be used to help a vision system to see, and to discover methods that make it possible to use such knowledge. How can one pursue this goal more effectively? There are several kinds of answers. The first is to abandon the restrictive format of line drawings, so that programs can use information about visual features that are not coded in this form. There is a large gain to be had by loosening up our formalisms to incorporate the many .pieces of special and general knowledge that are necessary for purposeful vision in the real world. Vision is primarily a.util tarian functmn in order to see properly it is necessary to know what kind of information is sought.</Paragraph> <Paragraph position="51"> We believe that progress in particular domains of recognition on the following problems will pay. large dividends: Getting Started: How to proceed from a set of feature clues to more global hypotheses. How to incorporate parallelism into this process: Keeping Going: How to recognize and mechanize confrontations between cohflicting hypotheses~ How to retain and use valid information that was incorporated in.</Paragraph> <Paragraph position="52"> hypotheses that have been refuted.</Paragraph> </Section> class="xml-element"></Paper>