AI Magazine Volume 11 Number 4 (1990) (© AAAI) BookReviews

Actors: A Model of in Simula (a language owed to 0. J. tational systems are seen as models of Concurrent Computation Dahl, B. Myhrhaug, and K. Nygaard social, economical, biological, and so and developed in the Norwegian on, organizations. The goal is to in Distributed Systems Computing Center in 1970) can be understand and build such computa- VarolAkman conceived of as ancestors of actors tional ecologies. Three issues of con- because they represent an individual cern in this regard are (1) the general Gul A. Agha’s Actors: A Model of Con- data item and all the operations and issues underlying open systems current Computation in Distributed Sys- procedures on it. However, actors are (Hewitt 1985), (2) the actual imple- tems (The MIT Press, Cambridge, Mass., more powerful than sequential pro- mentations of distributed computa- 1987, 144 pages, $25.00, ISBN O-262 cesses and data flow systems. One tion in computational ecologies, and 010925) is part of the MIT Press can represent a purely functional (3) the design of appropriate lan- Series in Artificial Intelligence. This system with an actor system. Similar- guages for open systems. Agha’s book volume is edited by Patrick Winston, ly, it is possible to specify arbitrary is a must for gaining an appreciation Michael Brady, and Daniel Bobrow. sequential processes with a suitable of the problems to be solved to In the actor formalism, pioneered actor system. However, the converse obtain these actual implementations by Carl Hewitt (1977), one perceives is not true. Although actor creation is and also deals with problems of abstract computational agents, called an integral part of the computational designing languages for open sys- actors, that are distributed in space. model of Hewitt and Agha, sequen- tems. If you want to understand the Each actor has a mailbox (and a mail tial processes such as communicating more practical ideas of ecological address) and associated with each sequential processes (CSP), owed to computation, theory of games, actor is a behavior. One actor can Tony Hoare, do not create other market economies and underlying influence the actions of another actor sequential processes (but can activate mechanisms, evolution and biologi- only by sending it a communication. other sequential processes). The cre- cal systems, and so on, there is a lot An actor can send another actor a ation of actors permits one to to gain in terms of conceptual clarity communication only when it knows increase the distributivity of compu- and computational fundamentals by the address of the second actor. An tation as it evolves. starting with this elegant book. Agha actor can (should be able to) receive Is programming a collection of also commented on this sociological more than one message simultane- actors easy? It is commonly accepted aspect of the actor model elsewhere: that parallel programming is, in gen- ously (in some sense). In case of syn- The term actor brings with it the eral, hard, and the need for yet chronous communication in nonactor image of an active entity, acting another programming language such paradigms, this simultaneous com- out its role in concert with as actors might seem difficult to munication requires the sender to others. Each actor can act only argue. One crucial observation here is wait until the recipient of the com- according to its script, which rep- that because message passing is fun- munication is ready to accept it. Once resents its view of the world. At damental to computation in actors, the recipient starts processing a given this level, the actor paradigm the time complexity of communica- communication, it is blocked from stands in sharp contrast to the tion might become the dominant accepting any other messages until logicist approach, which not factor in program execution. Thus, after it is done with the message-all only postulates the existence of a architectural considerations might for the sake of maintaining the integri- unique reality, but commits us to play a crucial role in the realistic ty of communications, no doubt. representing our knowledge in implementations of actor languages. The actor model achieves the same terms of a consistent collection Note, however, that the most attrac- effect through buffering. A recipient’s of information. mailbox is always ready to receive tive feature of actor languages is that messages. Furthermore, in the actor programmers are not required to The book has several additional framework, there are no restrictions code details such as when and where strengths that make it a useful source on freely communicating mail to use parallelism. Also, although for people working on distributed addresses; thus, the interconnections some functional programming lan- computation. Problems of distributed of the actors can be dynamically guages might have problems with computing such as divergence, dead- modified, and resource management history-sensitive shared objects, lock, and mutual exclusion are treat- is more flexibly done. actors can easily model such objects. ed. There is also an excellent chapter Should the reader begin pondering Recently, there has been interest- on abstraction and compositionality. the difference between actors and ing research into what is somewhat An excellent glossary of actor terms is objects (as found in object-oriented exotically termed the ecology of included as well as a list of references programming), I would like to make a computation (Hubermann 1988). In and a useful index. Although the book few clarifying observations. Objects this philosophy, distributed compu- is a revised version of the author’s

92 AI MAGAZINE BookReviews Readings from AI Magazine The First Five Years: 1980-1985 Edited with a Preface by Robert Engelmore AAAI is pleased to announce publication of Readings from AI Magazine, the complete collection of all the articles that appeared during AI Magazine’s first five years. Within this 650-page indexed volume, you will find articles on AI written by the foremost practitioners in the field-articles that earned AI Magazine the title “journal of record for the artificial intelligence community.” This collection of classics from the pre- mier publication devoted to the entire field of artificial intelligence is available in one large, paperbound desktop reference. SubjectsInclude:

q Automatic Programming n Distributed Artificial Intelligence n Games n Infrastructure n Learning n Natural Language Understanding n Problem Solving q Robotics n Education q General Artificial Intelligence n Knowledge Acquisition n Legal Issues n Object Oriented Programming n Programming Language n Simulation n Technology Transfer n Discovery w Expert Systems n Historical Perspectives n Knowledge Representation n Logic n Partial Evaluation n Reasoning with Uncertainty n Computer Architectures

$74 95 plus $2 postage and handling. 650 pages, illus., appendix, index ISBN O-929280-01-6. Send prepaid orders to The MIT Press, 55 Hayward Street, Cambridge, MA 02142. doctoral dissertation, it doesn’t carry physics, knowledge representation, and MIT Press, 1989, 266 pages, $25.00, the usual blemish associated with mathematical logic. ISBN O-262-12140-9), Dan Lloyd theses in general; that is, it is not seeks to provide answers that will written in the stream-of-conscious- References bridge the gap between computation- ness mode. al and connectionist models of the There is a well-known saying: Any Agha, G Foundational Issues in Concur- mind. He notes that naturalism theory that can be put in a nutshell rent Computing, Dept of Computer Sci- assumes that the human brain and ence, Yale Univ Unpublished. Also in the brains of some animals are belongs there. Well, running this Foundational Issues in Concurrent Com- risk, I’ll say that in a nutshell, the puting. In Proceedings of the ACM/- organs of thought or representation, actor approach is about the future of SIGPLAN Workshop on Object-Based but what is it about these brains that computing. Concurrent Programming, SIGPLAN makes them thoughtful? Additional- Postscript: A new (and probably Notices, eds. G. Agha, P. Wagner, and A. ly, just what sorts of physical systems the definitive) source has joined the Yonezawa. New York: Association of Com- are sufficient to embody the repre- ranks of actor literature: Concurrent puting Machinery. sentations typical of thinking? Final- Systems for Knowledge Processing: An Hewitt, C 1977. Viewing Control Struc- ly, could a nonbiological system-a Actor Perspective, edited by C. Hewitt tures as Patterns of Passing Messages Arti- computer perhaps-think? With and G. Agha (Cambridge, Mass., The ficial Intelligence 8(3): 323-364 these fundamental questions, Lloyd MIT Press, 1989). This book brings Hewitt, C 1985 The Challenge of Open introduces the central question to be together over 20 important contribu- Systems Byte lO(4): 223-242. answered by Simple Minds: How does tions on the actor concept and its Hubermann, B., ed. 1988. The Ecologyof the brain comprise the mind? applications. Although I haven’t seen Computation: Studies in Computer Science Lloyd seeks to outline a theory of this book, given my positive views and Artificial Intelligence, volume 2. The representing systems and lay the about the book under review, I believe Hague, The Netherlands: Elsevier. foundations of a natural philosophy that this new text will also be a plea- of mind. Such a reductionist theory, sure to read. he observes, must explain how repre- sentations are made out of nonrepre- Varol Akman is an assistant professor in Simple Minds sentational parts. Only by this means the Department of Computer Engineering can we solve the problem of mind and Information Science, Bilkent Univer- Lee A. Gladwin and brain and the open question of sity, Ankara, Turkey His current research concentrates on various theoretical aspects Of what are minds made? Internal the convergence of psychology and of AI, especially from the angle of intelli- mental representations? Matter? In neuroscience. gent computer-aided design systems, for this provocative and engaging work Simple Minds is divided into three example, commonsense reasoning, naive (Simple Minds, Cambridge, Mass.: The parts. Part 1, Representation from the

WINTER 1990 93 Book Reviews Bottom Up, describes what a theory noted with regard to neural networks: cal activity in the immense pharmacy of material representation must do local; distributed; and, what Lloyd of the brain is transformed by some and looks at two current approaches calls, featural. In local representation, alchemy into the language of thought, (internalism and externalism). It then each unit is dedicated to one con- but the reader is left far from seeing proceeds to develop a dialectic theory tent-understood as a proposition or how the hardware (or wetware) sup- of representation based on thought hypothesis-and the activation of ports the psychological software. experiments involving a simple device this unit indicates how committed A more orderly and better researched called Squint that integrates input the system is to the truth of the rep- bridging theory can be found in from multiple channels and can resentation. Distributed and featural Jean-Pierre Changeux’s Neuronal come to be a representative device. representation are nonlocal, with Man: The Biology of Mind (New York: Part 2, Interpreting Neural Networks, individual units participating in a Pantheon Books, 1985). Changeux, a applies the theory to biological and variety of representations. In dis- neurobiologist, provides a thorough simulated neural networks. Part 3, tributed representation, however, no and readable description of the From Simple Minds to Human Minds, single unit is exclusively associated brain’s physiology and then presents extends the theory to language use, with either a specific input or a spe- support for his own thesis that the consciousness, and reasoning. cific output, and all the active units brain contains representations of the Representation, according to Lloyd, must be interpreted together. Finally, outside world in the anatomical orga- depends on relations involving the in featural representation, every unit nization of its cortex and is capable world outside the representing system. is dedicated to the representation of a of building representations of its own He establishes three criteria for basic feature, and this interpretation and using them in its computations. representation: multiple channels, remains constant over time. Higher Much of Changeux’s description of convergence, and uptake. The first cognitive processes can thus be how multiple representations are requires the presence of multiple implemented by lower-level units. coordinated within the visual region information channels leading to the Again, there are clearer accounts of of the cortex, for example, would representation; the second requires feature detection, perception, and have proved useful to Lloyd in his the installation of an ‘and’ gate at the representation in Lynch and others. discussions of featural representation confluence of the channels to act as With this foundation, the reader is and metarepresentation. an indicator when conjunctions of prepared to cross over the metarepre- Similar comments could be made antecedent conditions occur. To work sentational bridge between the con- with reference to Donald 0. Hebb’s effectively, the convergence condi- nectionist and computational models cell assembly theory. First presented tion further requires that the multi- of the mind, which occurs in chapter in The Organization of Behavior (New ple-channel device be uniquely 6, “The Language of Thought.” First, York: John Wiley and Sons, 1949), it situated so that sometimes one event Lloyd summarizes Jerry Fodor’s was expanded and further document- (mediated by various channels) is suf- hypothesis that the inner representa- ed in Essay on Mind (Hillsdale, N.J.: ficient for its activation. In this way, tional system uses a language-like Lawrence Erlbaum Associates, 1980). channels converge on a single event. medium: Thought is the manipula- In the latter work, Hebb described The uptake criterion requires that the tion of internal symbols with cogni- the development of visual perception event representation have the capaci- tively relevant constituent structure. based on cell assemblies and sub- ty to cause either another representa- In particular, Lloyd examines the assemblies in the striate and peris- tion or a salient behavioral event. concept of mentalese, the uncon- trate cortex. His description of first- The introduction of these concepts scious, innate language of thought. and second-order assemblies accords does treble violence to Occam’s He then introduces his own con- well with Lloyd’s metarepresentation: Razor, adding unnecessary levels of cept of metarepresentation, a represen- When a group of assemblies are complexity to an already difficult tation inheriting content from repeatedly activated, simultane- topic. Most of the same ground is another representation. Lloyd offers ously or in sequence, those corti- better covered by Gary Lynch in the familiar example of newspaper cal cells that are regularly active Synapses, Circuits, and the Beginnings photos that are reconstructed from a following this primary activity of Memory (Cambridge, Mass.: The digitized representation of a photo- can themselves become orga- MIT Press, 1986), with far greater graph and concludes that what makes nized as a superordinate (second- attention to existing neurobiological representation is not the resemblance order) assembly. The activity of concepts and studies. There are also of features shared by representation this assembly then is representa- numerous fine illustrations in Lynch and metarepresentation but rather tive of the combination of to aid the reader in understanding the representation of original repre- events that individually orga- the neurological foundations of rep- sentational features by the metarepre- nized the first-order assemblies. resentation. sentation. Having set forth his dialectic theory Lloyd’s bridge struck me as promis- It is these higher-order cell assem- of representation in part 1, Lloyd ing in design but difficult to cross. blies that Hebb believed were the notes that his theory will illuminate Rather large leaps are required to get foundation of abstract thought. the human mind only if it can from one part of the bridge to the Although Lloyd’s metarepresenta- encompass the complexity of its sub- next. The reader jumps from a well- tional bridge is seemingly sound in ject. Bridging the gap between the described, easily visualized theory in design, it would have gained consid- brain, neural networks, and computa- part 1 to artificial-biological networks erable support from a broader read- tional models is the task of the in part 2 to metarepresentation in ing of the neurobiological literature remaining two parts of the book. part 3. There is scant but suggestive and the inclusion of references to the Three forms of representation are evidence provided that electrochemi- discoveries, concepts, models, and

94 Al MAGAZINE Book Reviews

theories already available. A new in his paper “Computing Machinery Goldkind uses elegant arguments to bridge can be built without inventing and Intelligence” (Mind 59, 1950). reject Malcolm’s conclusions. Never- new materials if existing ones are The author discusses the validity of theless, in examining an argument of already available. the Turing test and concludes that Malcolm’s, Goldkind fails to mention In some future golden age of sci- there are no means for showing that its lack of logical conclusiveness, ence, Lloyd notes, it might be possi- in contrast to a human being, a instead attacking its premises, which ble to test theories of the mind machine that passes the test is not is a bit confusing. The author closes against the details of the brain by able to think or understand. Thereby his last chapter with a repetition of directly observing inner representa- prepared, in the chapters that follow, the claim that a machine must be tions in their physical form. The the reader finds a more or less com- capable of knowledge to act purpose- results might vindicate a naturalism pact presentation of arguments fully, that is, to act at all. about the mind and show that the against the possibility of AI that have I found the critiques of Taylor’s representations posited by cognitive been established by philosophers in and Malcolm’s positions overly long psychologists or the beliefs and recent decades. and boring. It would have been better desires alluded to in folk psychology The book reviews and critiques the to shorten the discussion in these will turn out to be real, physical enti- major lines of argument in Dreyfus’s last two chapters and more deeply ties in the world. Alternatively, on book What Computers Can’t Do (New explore key terms in the critique. this last day when all mysteries will York: Harper & Row, 1972). To each In summary, this book discusses a be revealed, we might all stop and line of argument, the author responds couple of philosophical issues con- stammer, “So that’s how it works!” by pointing out logical gaps in Drey- cerning the possibility of AI. The first Lee A. Gladwin received his degree from fus’s argument. Although Goldkind two chapters, covering about one Carnegie-Mellon University. His disserta- shows that Dreyfus’s arguments do half of the book, provide an especial- tion was on historical problem solving. not prove the impossibility of AI, he ly good insight into common lines of He is currentlv interested in combining nevertheless emphasizes the value of argument against the possibility of hypermedia with neural networks. Dreyfus’s reasoning in giving pro- AI and their implicit suppositions. In found insight into the problems of each chapter, the author tries to constructing intelligent machines as weaken the arguments against the well as an understanding of the possibility of AI by pointing out cer- nature of human intelligence. tain unproven claims within each Machine Intelligence: A Goldkind then carries on a dialogue one. In making implicit assumptions Critique of Arguments about the possibility of machines explicit, the book certainly con- making mistakes, as raised in Turing’s tributes to the understanding of AI. against the Possibility of paper. Here, in contrast to the other Artificial Intelligence chapters, the author mainly follows Achim G. Hoffman is a member of the academic staff of the Computer Science Achim Homan his own thoughts. Goldkind’s case is Department at Technische Universitat that for a machine to err, it must Berlin. His current research interests are Stuart Goldkind’s book Machines and have intentions, which is the topic of machine learning and the application of Intelligence: A Critique of Arguments the remainder of the book. AI to very large-scale integrated design against the Possibility of Artificial Intel- Goldkind begins the examination automation. ligence (Westport, Conn.: Greenwood of this topic with a discussion of Press, 1987, 139 pages, $29.95, ISBN Richard Taylor’s 1966 monograph o-313-25450-8) considers and criti- Action and Purpose (Englewood Cliffs, cizes a collection of arguments N.J.: Prentice-Hall) in which Taylor against the possibility of AI. It does argues that machines are incapable of not claim to contain an exhaustive purposeful behavior. Goldkind points collection of such arguments or even out the restricted generality of to cover the most important argu- Taylor’s arguments to refute his line ments raised in the mainly philo- of thinking. Goldkind claims that sophical literature of the past few only machines that are capable of decades. Rather, as the author points knowledge are able to behave pur- out in the preface, the book is posefully. Unfortunately, Goldkind intended to examine the arguments does not exactly explain what he for and against the possibility of AI means by knowledge. As a result, the to give the reader more accurate reader can neither decide about the notions of terms such as intelligence conclusiveness of the author’s claim This publication and cognition as well as make the nor survey its implications. is available reader more aware of the implicit In the final chapter, parts of in microform suppositions involved in the debate Norman Malcolm’s 1968 article “The from University \ around the possibility of AI. Conceivability of Mechanism” (The Microfilms - 7 To pursue this issue, the book starts Philosophical Review, Vol. 77, pp. 45- International. by examining the well-known Turing 72) are reviewed. Malcolm argues for call Ml-free 600-521-3044.In Mtcbigan. Alaska and Hawaii call collect 313-761-4700 test as well as discussing some argu- an incompatibility of the causal mail inquiry to: Univenity Micmfii ments (for example, the argument of explanation of behavior on the one 300 North i&b Road. AM Arbx. MI 46106 consciousness) against the possibility hand and a purposive explanation of of AI that Turing himself addressed the same behavior on the other. For free information, circle no. 174

WINTER 1990 95