Early AI Research Articles

Towards Ranking Geometric Automated Theorem Provers

The field of geometric automated theorem provers has a long and rich history, from the early AI approaches of the 1960s, synthetic provers, to today algebraic and synthetic provers. The geometry automated deduction area differs from other areas by the strong connection between the axiomatic theories and its standard models. In many cases the geometric constructions are used to establish the theorems' statements, geometric constructions are, in some provers, used to conduct the proof, used as counter-examples to close some branches of the automatic proof. Synthetic geometry proofs are done using geometric properties, proofs that can have a visual counterpart in the supporting geometric construction. With the growing use of geometry automatic deduction tools as applications in other areas, e.g. in education, the need to evaluate them, using different criteria, is felt. Establishing a ranking among geometric automated theorem provers will be useful for the improvement of the current methods/implementations. Improvements could concern wider scope, better efficiency, proof readability and proof reliability. To achieve the goal of being able to compare geometric automated theorem provers a common test bench is needed: a common language to describe the geometric problems; a comprehensive repository of geometric problems and a set of quality measures.

Epochs of an AI cosmology

We began the discussion about the educational benefits to the general public of a standard AI Cosmology in Volume 4 Issue 2 of "AI Matters". But the mere notion of such a standard immediately raises lots of questions. In this column, we only focus on a very narrow set of those questions: First and foremost: Is there an AI Cosmology? If in fact the notion of cosmology applies to AI, is there only one valid AI Cosmology or are there competing valid cosmologies? If there is only one valid cosmology, what is its structure? How far can we take an AI Cosmology analogy? Like our Universe, did AI have the equivalent of a Big Bang? Is the field or structure of AI expanding or contracting? In an interview with Nikola Danaylov, podcaster of Singularity.FM, the late Marvin Minsky, one of the early AI pioneers, suggested that the evolution of AI in the 21st century maybe be contracting or even stalled! We posit that an AI Cosmology would help to educate the general public about our research and work, and its impact on society. Internally, it would help us to fix the objects, structure, and dynamics of the field. A standard cosmological model would help us all distinguish spurious AI mythos, pseudo science, hype, and fake news from the genuine article. A true cosmology of Artificial Intelligence would help us fix its origins, understand its trajectory, and predict its fate.

The computational therapeutic: exploring Weizenbaum\u2019s ELIZA as a history of the present

This paper explores the history of ELIZA, a computer programme approximating a Rogerian therapist, developed by Jospeh Weizenbaum at MIT in the 1970s, as an early AI experiment. ELIZA’s reception provoked Weizenbaum to re-appraise the relationship between ‘computer power and human reason’ and to attack the ‘powerful delusional thinking’ about computers and their intelligence that he understood to be widespread in the general public and also amongst experts. The root issue for Weizenbaum was whether human thought could be ‘entirely computable’ (reducible to logical formalism). This also provoked him to re-consider the nature of machine intelligence and to question the instantiation of its logics in the social world, which would come to operate, he said, as a ‘slow acting poison’. Exploring Weizenbaum’s 20th Century apostasy, in the light of ELIZA, illustrates ways in which contemporary anxieties and debates over machine smartness connect to earlier formations. In particular, this article argues that it is in its designation as a computational therapist that ELIZA is most significant today. ELIZA points towards a form of human–machine relationship now pervasive, a precursor of the ‘machinic therapeutic’ condition we find ourselves in, and thus speaks very directly to questions concerning modulation, autonomy, and the new behaviorism that are currently arising.

In Memoriam: Daniel G. Bobrow

Daniel G. Bobrow (1935–2017) was a research fellow at Xerox's Palo Alto Research Center and an early AI pioneer. Bobrow served as AAAI president from 1989–1991 and was a Fellow of AAAI and the Association of Computing Machinery (ACM). Prior to joining PARC, Bobrow was a vice president at BBN Laboratories, where we worked from 1965–1972. He won the ACM SIGOPS Hall of Fame Award for his contributions to the operating system TENEX. He also won the ACM Software System Award for his contributions to Interlisp. A celebration of his life was held on June 5 at the Palo Alto Research Center in Palo Alto, California.

The TOX-OER project: Learning Toxicology through Open Educational Resources. An International experience in permanent education

The objectives of this retrospective study were to use in-line milk progesterone (mP4) data to investigate relationships of (1) commencement of luteal activity (CLA), and (2) luteal phase (LP) length and frequency preceding first postpartum AI, with parity and AI outcomes in Canadian Holstein cows. Starting 21 ± 1 days postpartum (DPP), levels of mP4 were assessed every 2.2 ± 2.0 d through an automated in-line milk analysis system (Herd Navigator™, DeLaval International, Tumba, Sweden) until ∼55 d after first or second AI in 748 Holstein cows from two herds. The CLA was defined as the DPP of the first of at least two consecutive samples with mP4 ≥5 ng/mL, and the period with elevated mP4 (≥5 ng/mL) was defined as the LP. Cows were categorized by CLA [earlier (≤) or later (>) than 28, 35, 42, 49, 56, and 63 DPP], and by the pattern of LP frequency preceding first AI as having or not: (1) one or more normal LP (LP length ≥7 and ≤19 d); (2) one or more abnormal LP (LP length <7 or >19 d, or interluteal period ≥12 d); and (3) two or more LP (either normal or abnormal). Outcomes of first or second AI were determined by the interval between AI and cessation of the ensuing LP as: non-pregnant (mP4-decline ≤30 d), presumed-pregnant (no mP4-decline until 55 d), or presumed-pregnancy loss (mP4-decline between 31 and ≤55 d). The odds of pregnancy per AI (P/AI) at 55 d and pregnancy loss were evaluated using generalized linear mixed models. Primiparous cows had lower odds of having CLA ≤28 DPP [Odds ratio (OR) = 0.58, P = 0.002] and one or more abnormal LP (OR = 0.73, P = 0.04) than multiparous cows. In multiparous cows, CLA ≤28 DPP decreased pregnancy loss (OR = 0.48, P = 0.05) and CLA ≤56 DPP increased P/AI (OR = 4.69, P < 0.01) compared to a later CLA. Primiparous and multiparous cows that had one or more normal LP before first AI had increased P/AI (OR = 3.85 and 3.45, respectively, P < 0.01) and reduced pregnancy loss (OR = 0.26 and 0.27, respectively, P < 0.01) than cows without a normal LP. Primiparous cows that had one or more abnormal LP had decreased P/AI (OR = 0.62, P = 0.04) and increased pregnancy loss (OR = 1.64, P = 0.04) compared to those without an abnormal LP. In summary, AI outcomes were improved in multiparous cows that had early CLA and in cows of both parity groups that had at least one normal LP before first AI. However, primiparous cows that had at least one abnormal LP had reduced AI outcomes. Relationships between early postpartum luteal activity and AI outcomes were inconsistent between primiparous and multiparous cows.

Relationships among early postpartum luteal activity, parity, and insemination outcomes based on in-line milk progesterone profiles in Canadian Holstein cows

The objectives of this retrospective study were to use in-line milk progesterone (mP4) data to investigate relationships of (1) commencement of luteal activity (CLA), and (2) luteal phase (LP) length and frequency preceding first postpartum AI, with parity and AI outcomes in Canadian Holstein cows. Starting 21 ± 1 days postpartum (DPP), levels of mP4 were assessed every 2.2 ± 2.0 d through an automated in-line milk analysis system (Herd Navigator™, DeLaval International, Tumba, Sweden) until ∼55 d after first or second AI in 748 Holstein cows from two herds. The CLA was defined as the DPP of the first of at least two consecutive samples with mP4 ≥5 ng/mL, and the period with elevated mP4 (≥5 ng/mL) was defined as the LP. Cows were categorized by CLA [earlier (≤) or later (>) than 28, 35, 42, 49, 56, and 63 DPP], and by the pattern of LP frequency preceding first AI as having or not: (1) one or more normal LP (LP length ≥7 and ≤19 d); (2) one or more abnormal LP (LP length <7 or >19 d, or interluteal period ≥12 d); and (3) two or more LP (either normal or abnormal). Outcomes of first or second AI were determined by the interval between AI and cessation of the ensuing LP as: non-pregnant (mP4-decline ≤30 d), presumed-pregnant (no mP4-decline until 55 d), or presumed-pregnancy loss (mP4-decline between 31 and ≤55 d). The odds of pregnancy per AI (P/AI) at 55 d and pregnancy loss were evaluated using generalized linear mixed models. Primiparous cows had lower odds of having CLA ≤28 DPP [Odds ratio (OR) = 0.58, P = 0.002] and one or more abnormal LP (OR = 0.73, P = 0.04) than multiparous cows. In multiparous cows, CLA ≤28 DPP decreased pregnancy loss (OR = 0.48, P = 0.05) and CLA ≤56 DPP increased P/AI (OR = 4.69, P < 0.01) compared to a later CLA. Primiparous and multiparous cows that had one or more normal LP before first AI had increased P/AI (OR = 3.85 and 3.45, respectively, P < 0.01) and reduced pregnancy loss (OR = 0.26 and 0.27, respectively, P < 0.01) than cows without a normal LP. Primiparous cows that had one or more abnormal LP had decreased P/AI (OR = 0.62, P = 0.04) and increased pregnancy loss (OR = 1.64, P = 0.04) compared to those without an abnormal LP. In summary, AI outcomes were improved in multiparous cows that had early CLA and in cows of both parity groups that had at least one normal LP before first AI. However, primiparous cows that had at least one abnormal LP had reduced AI outcomes. Relationships between early postpartum luteal activity and AI outcomes were inconsistent between primiparous and multiparous cows.

AI and the Origins of the Functional Programming Language Style

The Lisp programming language is often described as the first functional programming language and also as an important early AI language. In the history of functional programming, however, it occupies a rather anomalous position, as the circumstances of its development do not fit well with the widely accepted view that functional languages have been developed through a theoretically-inspired project of deriving practical programming languages from the lambda calculus. This paper examines the origins of Lisp in the early AI programming work of the mid-to-late 1950s, and in particular in the work of Allen Newell, Cliff Shaw and Herbert Simon. Their 1956 program, the Logic Theory Machine, introduced new ideas about data and program structures that were articulated in response to perceived limitations in existing programming technique. Later writers, notably John Backus, have described these features as constituting a "programming language style" distinct from the traditional style that preceded it. The paper examines the origins of the earlier style in practices of manual computation, analyses the key technical differences between it and the style first manifested in the Logic Theory Machine, and concludes that programming practice and experience play a large and underappreciated role in the development of programming styles and languages.

Alcohol Expectancies Mediate the Relationship Between Age of First Intoxication and Drinking Outcomes in College Binge Drinkers

ABSTRACTBackground: While prior research has shown that age of first intoxication (AI) is associated with negative alcohol outcomes, limited research has examined factors accounting for this relationship. Alcohol expectancies, or beliefs about the effects of alcohol, may explain such associations as both positive and negative expectancies have been shown to be key predictors of drinking outcomes. Objective: The present study examined expectancies as mediators between early AI and alcohol-related outcomes. Method: Data collection occurred in 2012 and 2013. Participants were college students (N = 562, 65.8% women) who completed an online survey including measures of alcohol use history, alcohol expectancies, typical alcohol consumption, and alcohol-related problems. Structural equation modeling was used to test the hypothesized model. Results: Our findings support a model whereby AI is associated with drinking through its influence on both positive and negative expectancies. Specifically, an earlier AI was associated with stronger alcohol expectancies, which in turn, was associated with heavier alcohol use and alcohol-related problems. Conclusions/Importance: These findings are consistent with expectancy theory and previous research suggesting that more experienced drinkers hold stronger drinking-related beliefs, be it positive or negative, and these expectancies ultimately explain variability in alcohol use and problems. Our findings further support that expectancies play an important role in the initiation of drinking behavior.

Artificial Intelligence: From programs to solvers

Artificial Intelligence is a brain child of Alan Turing and his universal programmable computer. During the 60’s and 70’s, AI researchers used computers for exploring intuitions about intelligence and for writing programs displaying intelligent behavior. A significant change occurred however in the 80’s, as many AI researchers moved from the early AI paradigm of writing programs for ill-defined problems to writing solvers for well-defined mathematical models like Constraint Satisfaction Problems, Strips Planning, SAT, Bayesian Networks, Partially Observable Markov Decision Processes, and General Game Playing. Solvers are programs that take a compact description of a particular model instance and automatically compute its solution. Unlike the early AI programs, solvers are general as they must deal with any instance that fits the model. Many ideas have been advanced to address this crisp computational challenge from which a number of lessons can be drawn. In this paper, I revisit the problem of generality in AI, look at the way in which this ’Models and Solvers’ agenda addresses the problem, and discuss the relevance of this agenda to the grand AI goal of a computational account of intelligence and human cognition.

Biological, computational and robotic connections with Kant's theory of mathematical knowledge

In my research I meander through various disciplines, using fragments of AI that I regard as relevant, willing to learn from anyone whose ideas contribute. This makes me unfit to write the history of European collaboration on some area of AI research as originally intended for this collection. However, by interpreting the topic rather loosely, I can regard some European philosophers who were interested in Philosophy of mathematics as early AI researchers from whom I learnt much, such as Kant and Frege. Hume's work is also relevant. Moreover, more recent work by Annette Karmiloff-Smith, begun in Geneva with Piaget then developed independently, helps to identify important challenges for AI and theoretical neuroscience, that also connect with philosophy of mathematics and the future, rather than the history, of robotics. So this paper presents an idiosyncratic survey of a subset of AI stretching back in time, and deep into other disciplines, including philosophy, psychology and biology, and possibly also deep into the future.

Remarks on Receiving the Covey Award

My first remark on hearing that I was to receive the Covey Award (for “innovative research in the field of computing and philosophy”) was “Wow!”; the second, “That's great!”; and the third, “What a surprise!”. Perhaps those are not quite what the president of International Association for Computing and Philosophy (IACAP) had in mind when he offered to read “some remarks” on my behalf at the presentation ceremony. And certainly something more is needed here..... But they did express my delight at hearing of IACAP's recognition of my research. When I started my interdisciplinary work in this field, Hilary Putnam's paper on functionalism had just appeared (1960). But I didn't know about it. What I did know about were Alan Turing's provocative papers inMind (1950) and (even more exciting, I thought) in the Philosophical Transactions (1952); Kenneth Craik's little book on cerebral models (1943); Margaret Masterman's work on machine translation, which resolved ambiguities by using a thesaurus (1957, 1962); and a handful of early AI programs. These included the personality-model Argus (Reitman 1963; Reitman et al. 1964), an intriguing computer model of neurosis (Colby 1963, 1964; Colby and Gilbert 1964), a model of emotionally biassed cognition (Abelson 1963), the draughts player that learnt to beat its programmer (Samuel 1959), and the General Problem Solver (Newell and Simon 1961). Last, but certainly not least, I had been spurred by George Miller's intoxicating manifesto for cognitive science, Plans and the Structure of Behavior (Miller et al. 1960). All of these related to longstanding philosophical questions that had bothered me since my high school days. They had led me to read medicine (intending a career in psychiatry) and also philosophy at Cambridge—where Masterman was an inspiring, if often infuriating, supervisor. They concerned the mind–body problem in general, and the nature of human freedom and psychopathology in particular. As an undergraduate, I had published a paper inMind that addressed such questions—but in a traditional manner (Boden 1959). For despite having found both the papers of Turing hugely exciting, and theMTwork highly ingenious, I couldn't yet see how to apply them to Philos. Technol. (2013) 26:333–339 DOI 10.1007/s13347-013-0115-x

Oliver G. Selfridge (1926-2008)

Oliver G. Selfridge, a leader in early artificial intelligence research, whose work extended from 1956 to 2008, recently passed away. His research in machine learning, neural networks and agent-based computing still influence AI today. He helped organize the 1956 Dartmouth summer school, which is generally seen as the origin of American AI and developed Pandemonium, an early AI program that dealt with learning and reasoning, during his more than 50-year career.

Influence of stage of lactation and milk production on conception rates after timed artificial insemination following Ovsynch

Conception rates after timed artificial insemination (TAI) are of paramount importance for the success of protocols based on synchronization of ovulation. Stage of lactation and milk production level are known factors that influence dairy cow fertility. It was the objective of this study to analyse the effect of stage of lactation and milk production level on conception rates and pregnancy rates by 200 days in milk (DIM) in dairy cows synchronized with the Ovsynch protocol (Day −10, Day −1: 0.1 mg of D-Phe6-gonadorelin, Day −3: 0.5 mg of cloprostenol, Day 0: AI). A total of 1288 dairy cows were assigned to two groups and classified in three production levels (high, average, low). Cows of all milk production levels in Group 1 (Simultaneous Ovsynch, SO) were synchronized with the Ovsynch protocol simultaneously for TAI between 73 and 81 DIM. In Group 2 cows with average milk production were synchronized at the same time as Group 1, while low producing cows were synchronized 3 weeks earlier and high producing cows were synchronized 3 weeks later than Group 1, respectively. First service conception rates (FSCRs) were lower ( P<0.05) in cows synchronized earlier than in cows of the same production level synchronized later (low production: 14.4% (22/153) versus 34.5% (51/148); high production: 28.2% (40/142) versus 41.4% (53/128)). Milk production level had no significant impact on conception rates after TAI in cows synchronized at the same stage of lactation. At 200 DIM fewer cows with high production level were pregnant than cows with average or low production ( P<0.05). This effect was independent of the stage of lactation at the initiation of Ovsynch. Endometritis at a postpartum examination did not influence conception rates after TAI. In conclusion, stage of lactation, but not milk production level, has a major influence on conception rates after TAI. Early AI after Ovsynch is less efficient and therefore its return on investment should be evaluated carefully.

Achiasmatic male meiosis in Cimex sp. (Heteroptera, Cimicidae)

Chromosomes and their behaviour in spermatogenesis in Cimex species (Heteroptera, Cimicidae) from a bat species Myotis emarginatus were studied by using both a conventional staining method and sequence-specific fluorochromes CMA 3 and DAPI. Spermatogonial metaphase revealed 28 auto- somes, one large Y chromosome and four small X chromosomes. The Y was almost fully C-band positive as well as most telomeres of autosomes. The same regions appeared to be CMA 3 and DAPI positive, indicating that in hete- rochromatin GC-rich clusters are dispersed within AT-rich repeats. In meiosis homologous chromosomes in bivalents are physically associated with one or sel- dom two sites or collochores, i.e. meiosis is achiasmatic. In MI cells chromosomes in bivalents were aligned in parallel, but at early AI bivalents open out and chro- mosomes move to the poles one telomere foremost. Meiosis is post-reductional for the sex chromosomes.

As advertised: A review of The MIT Encyclopedia of the Cognitive Sciences

For me, this is an exciting time to be a cognitive scientist and a cognitive linguist. Cognitive Linguistics has developed rapidly and with enormous success over the past two decades, providing a cognitively based account of language. When results in cognitive linguistics are taken together with results in the other cognitive sciences, a radically new view of the mind and language—and their relation to the brain—emerges. As a result, the original formalist nativist paradigm of cognitive science as it developed in the 1960s and early 1970s has been stood on its head. I was one of the originators of that paradigm, among the researchers first bringing formal logic as an account of natural language semantics into linguistics in the early 1960s. The hope then was to fit logic and Chomskyan transformational generative grammar into a unified approach to language and mind. The formalist nativist paradigm that subsequently developed tried to fulfill that dream, with the hope of merging Anglo-American analytic philosophy with formal logic, generative grammar, early AI, cognitive psychology, and cognitive anthropology. By the mid-1970s, it was clear that the formalist nativist paradigm did not fit the facts. Research by Brown, Berlin, Kay, Slobin, Rosch, Mervis, Barsalou, DeValois, McNeill, and others indicated that words and concepts not only did not fit formal logic, but were fundamentally embodied and connected to human experience, with very different properties. For example, basic color terms in all the world’s languages show commonalities derived from the neurophysiology of color vision. More generally, words and concepts show gradations, prototype structures of at least half-a-dozen kinds, radial structure, and basic-level structure deriving from many aspects of direct experience. These results and many others led to the development of cognitive linguistics and an embodied cognitive science beginning in the late 1970s.