Analogy and Qualitative Representations in the Companion Cognitive Architecture

Articles

Kenneth D. Forbus, Thomas Hinrichs

n The Companion cognitive architec - very cognitive architecture starts with a set of theoreti - ture is aimed at reaching human-level cal commitments. We have argued (Forbus 2016) that AI by creating software social organ - Ehuman-level artificial intelligences will be built by cre - isms — systems that interact with peo - ating sufficiently smart software social organisms. By that we ple using natural modalities, working mean systems capable of interacting with people using natu - and learning over extended periods of ral modalities, operating and learning over extended periods time as collaborators rather than tools. Our two central hypotheses about how of time, as apprentices and collaborators, instead of as tools. to achieve this are (1) analogical rea - Just as we cannot directly access the internal representations soning and learning are central to cog - of the people and animals we work with, cognitive systems nition, and (2) qualitative representa - should be able to work with us on our terms. But how does tions provide a level of description that one create such systems? We have two core hypotheses, facilitates reasoning, learning, and inspired by research in cognitive science: communication. This article discusses Our first core hypothesis is that analogical reasoning and the evidence we have gathered support - learning are central to human cognition. There is evidence that ing these hypotheses from our experi - processes described by Gentner’s (1983) structure-mapping ments with the Companion architec - ture. Although we are far from our theory of analogy and similarity operate throughout human ultimate goals, these experiments pro - cognition, including visual perception (Sagi, Gentner, and vide strong evidence for the utility of Lovett 2012), reasoning and decision making (Markman and analogy and qualitative representation Medin 2002), and conceptual change (Gentner et al. 1997). across a range of tasks. We also discuss Our second core hypothesis is that qualitative representations three lessons learned and highlight three (QRs) are a key building block of human conceptual structure. Con - important open problems for cognitive tinuous phenomena and systems permeate our environment systems research more broadly. and our ways of thinking about it. This includes the physical world, where qualitative representations have a long track record of providing human-level reasoning and performance (Forbus 2014), but also in social reasoning (for example, degrees of blame [Tomai and Forbus 2007]). Qualitative rep - resentations carve up continuous phenomena into symbolic descriptions that serve as a bridge between perception and cognition, facilitate everyday reasoning and communication, and help ground expert reasoning.

The focus of the Companion cognitive architec - roles in cognition. We close with some lessons ture (Forbus, Klenk, and Hinrichs 2009) is on high - learned and open problems. er-order cognition: conceptual reasoning and learn - ing, and learning through interactions with others, by contrast with architectures that have focused on Analogical Reasoning and Learning skill learning (for example, ACT-R [Anderson 2009] Dedre Gentner’s (1983) structure-mapping theory and SOAR [Laird 2012]). In Newell’s (1990) time- proposed that analogy involves the construction of scale decomposition of cognitive phenomena, con - mappings between two structured, relational repre - ceptual reasoning and learning occur in what are sentations. These mappings contain correspondences called the rational and social bands ,1 unlike many (that is, what goes with what), candidate inferences architectures, which start with Newell’s cognitive (that is, what can be projected from one description band (Laird, Lebiere, and Rosenbloom 2017). Thus to the other, based on the correspondences), and a we approximate subsystems whose operations occur score indicating the overall quality of the match. at faster time scales, using methods whose outputs Typically, the base is the description about which have reasonable cognitive fidelity, although we are more is known, and the target is the description about not making theoretical claims about them. For which one is trying to reason about, and hence infer - example, in Companions constraint checking and ences are made from base to target by default. Infer - simple inferences are carried out through a logic- ences in the other direction, reverse candidate infer - based truth-maintenance system (Forbus and de ences, can also be drawn, which is important for Kleer 1994). Similarly, the Companion architecture difference detection, discussed more later. is implemented as a multiagent system, capable of There is ample evidence for the psychological plau - running on a single laptop or distributed across sibility of structure mapping, out of Gentner’s lab many cluster nodes, depending on the task. We sus - and others (Forbus 2001). This evidence includes pect that many data-parallel operations and a num - some implications that may surprise AI researchers. ber of coarse-grained parallel operations are impor - Here are three such findings: tant for creating robust software organisms, but this (1) The same computations that people use for current organization is motivated more by engineer - analogy are also used for everyday similarity (Mark - ing concerns. By contrast, we spend considerable man and Gentner 1993). Most AI researchers model effort on sketch understanding and natural language similarity as the dot product of feature vectors or oth - understanding, since interaction through natural er distributed representations. But those approaches modalities is a core concern. are not compatible with psychological evidence that Cognitive systems are knowledge rich (for exam - indicates, even for visual stimuli, that relational rep - ple, McShane [2017]). The Companion architecture resentations are important (Goldstone, Medin, and uses the Cyc 2 ontology and knowledge base (KB) con - Gentner 1991). tents, plus our own extensions including additional (2) Differences are closely tied to similarity (Mark - linguistic resources, representations for visual/spatial man and Gentner 1996). There is an interesting dis - reasoning, and the hierarchical task network (HTN) sociation between difference detection and naming a plans that drive Companion operations. This is in difference: it is faster to detect that things are differ - sharp distinction with current practice in machine ent when they are very different, but faster to name learning, where the goal is that systems for every task a difference when they are very similar (Sagi, Gen - must be learned from scratch. The machine-learning tner, and Lovett 2012). This falls directly out of our (ML) approach can be fine for specific applications, computational model of analogical matching below. but it means that many independently learnable fac - (3) Analogical comparison, especially within- tors in a complex task must be learned at the same domain comparison, can happen unconsciously. For time. This intermingling leads to requiring far more example, people reading a story can be pushed training data than people need, and reduces transfer between two different interpretations of it, based on of learned knowledge from one task to another. The a different story that they read previously (Day and cognitive-systems approach strives for cumulative Gentner 2007). Moreover, they report being sure learning, where an architecture that is general pur - that the second story was complete in itself, and pose can learn to do one task and successfully applies nothing from the previous story influenced them. that knowledge to learn other tasks. This implies, for example, that in any of the System The remainder of this article summarizes the evi - 1 / System 2 models (for example, Kahneman dence provided by Companion experiments for our [2011]), analogy and similarity are actually used in core hypotheses. We start by summarizing the ideas both systems. underlying our hypotheses, briefly introducing struc - Our computational models of analogical processes ture mapping and our models of its processes for have been designed both as cognitive models and as analogy followed by the key aspects of qualitative performance systems. That is, each of them has been reasoning. Then we describe the evidence that anal - used to explain (and sometimes predict) psychologi - ogy and qualitative representations play a variety of cal findings, while also being used in AI systems

WINTER 2017 35 Articles

whose sole goal was achieving some new functional - world, we must deal with quantities like heat, tem - ity. These models are SME, MAC/FAC, and SAGE. perature, and mass, and arrange for processes like The structure-mapping engine (SME) (Forbus et al. flows and phase changes to cook. Spatial factors are 2016) models analogical mapping. It computes up to key in many physical situations, such as finding safe three mappings, using a greedy algorithm to operate ground during a flood. In social reasoning, we think in polynomial time. It can also operate incremental - about quantities like degree of blame, how strong a ly, extending mappings as new facts are added to the friendship is, and the relative morality of choices we base or target, which can be useful in problem solv - must make. Even in metacognition, we estimate how ing and language understanding. difficult problems are, and decide whether to carry Many are called but few are chosen (MAC/FAC) on or give up on a tough problem. Qualitative repre - (Forbus, Gentner, and Law 1995). MAC/FAC models sentations provide more abstract descriptions of con - analogical retrieval, using two map/reduce stages. tinuous parameters and phenomena, which are more The MAC stage uses a nonstructured vector represen - in line with data actually available to organisms. We tation, automatically derived from structured repre - may not be able to estimate accurately how fast sentations, such that their dot products provide an something is moving, but we can easily distinguish estimate of how SME will score the original structur - between rising and falling, for example. People rou - al representations. The best, and up to two more if tinely come to reasonable conclusions with very lit - sufficiently close, are passed to the FAC stage, which tle information, and often do so rapidly. Combining uses SME in parallel to match the original structured qualitative representations with analogy provides an representations. The MAC stage provides scalability, explanation for this (Forbus and Gentner 1997). while the FAC stage provides the precision and infer - Broadly, qualitative representations can be decom - ences needed to support reasoning. posed into those concerned with quantities and The sequential analogical generalization engine those concerned with space. Quantities are an impor - (SAGE) (McLure, Friedman, and Forbus 2015) models tant special case because many continuous parame - analogical generalization. For each concept, it main - ters are one dimensional, and this imposes addition - tains a generalization pool that is incrementally al constraints on reasoning. We use qualitative updated as new examples arrive. The pool contains process theory (QP theory, Forbus [1984]) as our generalizations, which are structured descriptions account of quantities and continuous processes. Spa - with a probability attached to each statement, and tial qualitative representations are typically ground - unassimilated examples. For each new example, the ed in metric representations (that is, the metric dia - most similar item from the pool is retrieved through gram/place vocabulary model [Forbus, Nielsen, and MAC/FAC. If sufficiently similar, they are assimilat - Faltings 1991]), which provide the functional equiv - ed. The assimilation process involves updating the alent of visual processing to support spatial reason - probabilities for each statement, based on the con - ing. Once places are constructed, often purely quali - tents of the match and the differences, and replacing tative reasoning can be done on them, through any nonidentical entities with skolems. Thus com - spatial calculi (Cohn and Renz 2008). Our current monalities become highlighted and accidental infor - model of visual and spatial reasoning, CogSketch mation becomes deprecated. Note that logical vari - (Forbus et al. 2011), automatically produces cogni - ables are not introduced, since generalizations can tively plausible visual and spatial representations still be applied to new situations through analogy. from vector inputs (for example, digital ink, In computational terms, we think of these as an copy/paste from PowerPoint), and is integrated into analogy stack: An organization of analogical process - the Companion architecture. ing for cognitive systems that we believe has wide applicability. In the Companion cognitive architec - Some Roles of Analogy and Qualita - ture, analogical operations are more primitive than back chaining. Rule-based reasoning and logical tive Representations in Cognition deduction are used for encoding, case construction, We provide evidence for our hypotheses by examin - constraint checking, and simple inferences. The ing a range of tasks, showing how analogy and qual - architecture also includes an AND/OR graph problem itative representations have been used to enable the solver and an HTN planner, but all of these freely use Companion architecture to perform them. 3 the analogical processes above in their operations. Reasoning Qualitative Representations Textbook problem solving involves solving the kinds of problems that are posed to students learning sci - People have robust mental models that help them ence and engineering. In experiments conducted operate in the world; for example, we cook, navigate, with the aid of the Educational Testing Service (ETS), and bond with others. A common concern in these we showed that a Companion, using representations models is how to deal with continuous properties generated by ETS and Cycorp, could learn to transfer and phenomena. In reasoning about the physical knowledge through analogy to new problems across

36 AI MAGAZINE Articles six different within-domain conditions (Klenk and tions for observed behavior (Forbus 2015). Moreover, Forbus 2009). It operated by using MAC/FAC to even partial explanations relating aspects of a situa - retrieve worked solutions, which were at the level of tion can be projected through candidate inferences, detail found in textbooks, rather than the internals of without a complete and correct logically quantified the reasoning system, which neither ETS nor Cycorp domain theory. For example, questions from a test of had access to. Equations to solve new problems were plausible reasoning can be answered by chaining mul - projected as candidate inferences onto the new prob - tiple small analogies, similar to how rules are chained lem, with their relevance ascertained by overlapping in deductive reasoning (Blass and Forbus 2016). event structure between the new and old problem. This included solving larger, more complex problems Understanding Language by combining analogies involving two solutions from Our research indicates that analogy can play at least simpler problems. Qualitative representations played four roles in natural language understanding. First, three roles in this reasoning: First, some problems analogies are often explicitly stated in language, par - themselves are qualitative in their nature, such as ask - ticularly in instructional materials. These need to be ing whether something will move at all in a given sit - detected so that subsequent processing will gather up uation, and if so, will its velocity and acceleration be the base and target and the appropriate conclusions constant or increasing. The second role was express - can be drawn. Analogical dialogue acts (Barbella and ing the conditions under which an equation is appli - Forbus 2011) extend communication act theories cable. The third role was translating real-world con - with additional operations inspired by structure map - ditions into parameter values (for example, at the top ping, enabling a Companion to learn from such of its trajectory, the velocity of a rising object is zero). analogies expressed in natural language, as measured The role of qualitative representations for reason - by improved query-answering performance. Second, ing found in earlier stages of science education is for integrating knowledge learned by reading, it can even larger. Before high school, children are not asked be useful to carve up material in a complex text into to formulate equations, but to reason qualitatively, cases. These cases can be used for answering compar - draw conclusions from graphs and tables, and tie ison questions directly (Barbella and Forbus 2015), knowledge learned in class to their everyday experi - and for subsequent processing (for example, rumina - ence. An analysis of fourth-grade science tests sug - tion [Forbus et al. 2007]). Third, analogy can also be gests that qualitative representations and reasoning used to learn evidence for semantic disambiguation, are sufficient for at least 29 percent of their questions by accumulating cases about how particular (Crouse and Forbus 2016), and even more when qual - word/word sense pairings were used in the linguistic itative spatial and visual representations are taken context (including lexical, syntactic, and semantic into account. information), and honed by SAGE to suggest how to disambiguate future texts (Barbella and Forbus 2013). Decision Making Finally, analogy can be used to learn new linguistic An example of decision making is MoralDM (Deghani constructions (McFate and Forbus 2016a), for exam - et al. 2008), which models the influence of utilitarian ple, understanding sentences like “Sue crutched Joe concerns and sacred values on human decision mak - the apple so he wouldn’t starve.” ing. An example of a sacred value is not taking an Our evidence to date suggests that QP theory pro - action that will directly result in someone’s death, vides a formalism for natural language semantics con - even if it would save more lives (see Scheutz [2017] for cerning continuous phenomena. Qualitative repre - examples). MoralDM modeled the impact of sacred sentations capture the level of detail found in values through a qualitative order of magnitude rep - linguistic descriptions, and mappings can be estab - resentation, so that when a sacred value was involved, lished from FrameNet representations (McFate and the utilitarian differences seemed negligible compared Forbus 2016b) to qualitative process theory. More - to violating that value. While it had some rules for over, qualitative representations can be learned by detecting that a situation should invoke a sacred val - reading, with the knowledge produced improving ue, it also used analogy to apply such judgements strategy game performance (McFate, Forbus, and Hin - from similar situations. Experiments indicate that richs 2014). increasing the size of the case library improved its decision, as measured against human judgments in Domain Learning independent experiments (Deghani et al. 2008), and Learning from instruction often involves multiple using analogical generalization over cases improved modalities, for example, text and diagrams. Analogy performance further (Blass and Forbus 2015). can be used to fuse cross-modal information, by using SME to align the representations constructed for each Commonsense Reasoning modality and projecting information across them. For Analogical reasoning may be at the heart of com - example, Lockwood (Lockwood and Forbus 2009) monsense reasoning: Similar prior experiences can be showed that semiautomatic language processing of a used to make predictions but also provide explana - simplified English chapter of a book on basic

WINTER 2017 37 Articles

machines, combined with sketches for the diagrams, successfully modeled the trajectories of intuitive enabled a system to learn enough to correctly answer force models (Friedman and Forbus 2010) using 12 out of 15 questions from the back of the chapter. sketched behaviors, transitions between misconcep - Chang (2016) has shown that these ideas can work tions when learning about the human circulatory when the language processing is fully automatic. By system (Friedman and Forbus 2011), misconceptions using a corpus of instructional analogies aimed at mid - about why there are seasons (Friedman, Forbus, and dle-school instruction, expressed in simplified English Sherin 2011), and debugging misconceptions about and sketches, a Companion learned enough to cor - the day/night cycle by analogy (Friedman, Barbella, rectly answer 11 out of 14 questions from the New and Forbus 2012). York Regent’s examination on the topics covered. Companions have also been used to do cross- Cognitive State Sensing domain analogical learning, both in physical Our mental life is governed in part by the ability to domains (Klenk and Forbus 2013) and in games (Hin - sense our internal state: in doing a forced-choice task, richs and Forbus 2011). For cross-domain learning, like the one in figure 1, often one choice “looks building up persistent mappings across multiple right” and we take it, even if we cannot always artic - analogies, essentially making a translation table ulate why this is so. If it isn’t obvious, we have to between domains, was important. In cross-domain think about it more. We believe that the structural game learning, SME was used to build up a metamap - evaluation score computed by SME is a signal used in ping between predicates in the domains, which then these computations. In simulating this task (Kan - enabled the rest of the domain knowledge to be daswamy, Forbus, and Gentner 2014), if there is a imported. strong difference between SME’s scores for the two Qualitative representations provide a useful level comparison, the highest score is chosen. But if not, of causal knowledge for learning the dynamics of rerepresentation techniques are used to modify the complex domains (Hinrichs and Forbus 2012a), and encodings of the stimuli (which are automatically can be used as a high-level representation for goals generated through CogSketch) to attempt to discrim - and strategies (Hinrichs and Forbus 2014). Qualita - inate between them. This model has captured the tive reasoning supports reflective reasoning by treat - phenomena in several psychological experiments. ing goal activation levels as continuous quantities in Emotions are another kind of cognitive state sens - order to represent mental states (Hinrichs and Forbus ing, in our view. One property of emotions is that 2016). For example, in reasoning about a strategy they are often rapid, but not always specific. Wilson, game such as Freeciv, 4 the learned qualitative model Forbus, and McLure (2013) suspect this is because of domain dynamics enables identifying trade-offs, appraisal information used in generating emotions is and high-level strategies such as first expanding one’s stored with memories and retrieved through analogy empire, then building up its economy, can be to produce a rapid response. This response is subse - expressed by processes that are carried forward by a quently modified by slower cognitive processes, and player’s actions. eventually when new memories are consolidated, the appraisals are adjusted based on the outcome of the Conceptual Change situation. Hence a problem that might have initially One of the well-documented phenomena in cogni - seemed scary, if mastered, gets stored as something tive science is conceptual change, that is, how peo - doable. With this multiphase model of emotions, a ple’s mental models develop over long spans of time Companion learning to solve problems was able to (for example, diSessa, Gillespie, and Esterly [2004]; perform more effectively, and with emotional Ioannides and Vosniadou [2002]). Catalogs of mis - dynamics consistent with the literature. conceptions have been created for a variety of phe - nomena, and in some cases, trajectories that learners Classification take in mastering these concepts have been analyzed The analogy stack described earlier can be used for and documented. Friedman’s assembled coherence classification, by doing analogical retrieval over the theory of conceptual change (Friedman 2012) takes union of generalization pools representing the possi - as its starting point qualitative, compositional mod - ble concepts of interest. The label associated with the el fragments, but initially tied to specific experiences. pool where the best reminding came from serves as Analogical retrieval is used to determine which mod - the identification of the concept, with the score pro - el fragments are retrieved and combined to create a viding an estimate of the quality of the classification model for a new situation. This theory explains both and the correspondences of the mapping providing why mental models often seem very specific to par - an explanation as to why that classification is appro - ticular classes of situations and yet, within those priate. The candidate inferences provide further sur - classes, are relentlessly consistent. Model fragments mises about the new example. This method has been are constructed through SAGE and a set of transfor - used to learn geospatial concepts in a strategy game mation heuristics, and applied through MAC/FAC. (for example, isthmus, bay, and others), achieving 62 Friedman’s Companion-based TIMBER system has percent accuracy. By extending SAGE with automat -

38 AI MAGAZINE Articles

Figure 1. An Example of a Forced Choice Task. ic near-miss detection, accuracy rose to 77 percent become task specific instead of general purpose. We (McLure, Friedman, and Forbus 2015). believe this is a very useful approach that can be applied to many techniques, for example, using SME A Meta-Analogy as a kernel for SVMs. But we know of no evidence that Analogy is more conservative than explanation-based this is psychologically plausible. learning, which generalizes after only one example, but requires many fewer examples than feature-based statistical learning systems — sometimes orders of Some Lessons Learned magnitude fewer (for example, Kuehne, Gentner, and We have learned a lot through these and other exper - Forbus [2000]). Can we do even better by somehow iments. We summarize the three most important les - combining ideas of structure-mapping with ideas from sons here. machine learning? Here is a meta-analogy that sug - First, the models of structure-mapping processes gests a family of approaches: structure mapping is to turn out to be surprisingly robust. We have only relational representations what dot product is to fea - made two extensions in tackling the span of tasks we ture vectors. This suggests that any machine-learning have discussed. The first are interim generalizations, technique defined on feature vectors originally might that is, a small set of generalization pools that exist in be adapted to a hybrid method by replacing dot prod - working memory, in contrast with normal general - uct with SME. For example, by combining a structured ization pools, which are held in long-term memory. form of logistic regression with SAGE on the link plau - Interim generalizations were motivated by both the sibility/knowledge base completion task, Liang and rapid learning of children in experiments and by the Forbus (2015) showed that this approach can yield unconscious story intrusion effects in adults. In the state-of-the-art results, with several orders of magni - experiments listed, they are used in learning forced- tude fewer examples, while also providing material for choice tasks as a simple form of rerepresentation, that explanations. Another hybrid has been shown to pro - is, when an interim generalization is retrieved for an duce state of the art performance on the Microsoft Par - item, nonoverlapping aspects of the item are filtered aphrase task (Liang et al. 2016). The advantage of using out. The second are near misses. Winston’s (1970) task-specific training is that similarity weights can be concept of near miss required a teacher to know the tuned to specific tasks, and learned negative weights machine’s single concept, and formulate an example (which SME does not produce) can improve discrimi - that differed in only one way. McLure’s observation nation. The drawback is that the representations (McLure, Friedman, and Forbus 2015) was that learned, unless the weights are stored separately, retrieving an item from a different generalization

WINTER 2017 39 Articles pool over items in the intended label Every operation is done with respect to our analogical models through hard - is, when the two labels are mutually a logical environment, for example, a ware, to reach this scale and maintain exclusive, a near miss. He extended microtheory and those it inherits from, real-time performance. The second is SAGE to construct positive and nega - thereby controlling what knowledge is that being able to use natural modali - tive probabilistic hypotheses based on used in a computation. Alternatives and ties to interact with people, in a fluent near misses, which are added to the qualitative states are represented as and adaptive manner, is a key bottle - retrieval criteria for the generalization microtheories, as are subsketches with - neck for progress. While for Compan - pool to improve discriminability. in a sketch. In our reasoning engine, ions simplified English syntax has Second, type-level qualitative repre - microtheories are also cases, which been very productive, we would prefer sentations (Hinrichs and Forbus makes analogical reasoning even more them to be able to puzzle through 2012b) are important for scalability tightly integrated. more complex syntax. Our current and for flexible reasoning. In tradition - hope is that learning linguistic con - al qualitative reasoning, a system structions through analogy will enable might start with a schematic or sce - Discussion and this, but that is far from clear. The final nario description, and then formulate Open Problems issue is longevity, that is, creating cog - and use models to reason about it. By nitive systems that successfully per - Our experience with Companions to form while learning over weeks, contrast, consider playing a strategy date has provided strong evidence for game like Freeciv, where constructing a months, and years at a time. Most cog - our core hypotheses. Analogical pro - nitive architectures, including ours, are civilization is the goal. A successful civ - cessing and qualitative representations ilization involves a dozen or more fired up for experiments, the experi - are useful for textbook problem solv - ments are run, and then they are shut cities, each with many parameters, and ing, moral decision making, and com - even more units. Formulating and down. Achieving robust learning while monsense reasoning. Analogy can play maintaining accurate performance, maintaining explicit propositional multiple roles in natural language qualitative models for each city, much even with people inspecting the sys - understanding and visual problem tem’s internals, is quite challenging less building a complete snapshot of solving, and qualitative representa - the game state, can be prohibitively (Mitchell et al. 2015). Enabling Com - tions provide a natural level of expres - panions to build up solid internal expensive. Moreover, it is often neces - siveness for semantics and reasoning. sary to reason about planned units and models of their own reasoning and Together they can successfully model a learning processes seems to be the cities that don’t exist yet. Type-level variety of phenomena in conceptual representations make these problems most promising approach to us. We change, and provide a means of fusing think that the notions of attention and simpler, although still not easy. They information across modalities. They control that Bello and Bridewell (2017) are amenable to analogy (because they can be used inside cognitive systems explore will be an important part of are quantifier free) and also useful in for cognitive state sensing, to improve the solution to longevity. natural language semantics: The QP problem solving. They can be used for frame representations we generate are classification tasks. And there is now Acknowledgments neutral with respect to whether they evidence that, by combining analogi - We thank the Office of Naval Research, are specific or generic, leaving this cal processing with ideas from stan - the Air Force Office of Scientific decision to later semantic processing, dard machine learning, state of the art Research, the National Science Foun - which has the contextual information performance on tasks of interest to the dation, and the Defense Advanced needed to determine this. machine learning community can be Research Projects Agency for their sup - Finally, we have found the Cyc ontol - attained, sometimes with orders of port through multiple programs. State - ogy and KB contents to be extremely magnitude fewer examples and in - ments and opinions expressed may not valuable. We developed our own rea - creased explainability. reflect the position or policy of the soning engine because we needed To be sure, many open problems United States government, and no offi - source code access and, for us, analogy remain. We outline three here. The cial endorsement should be inferred. operations are central. The breadth and first is scaling. How much knowledge David Barbella, Maria Chang, Morteza depth of the Cyc ontology has saved us might be needed to achieve human- Dehghani, Scott Friedman, Matthew countless hours, by letting us use exist - level AI? Aside from the initial endow - Klenk, Kate Lockwood, Andrew Lovett, ing representations instead of building ment of knowledge, how much experi - Joe Blass, Max Crouse, Subu Kan - our own. Even when our requirements ence might be needed? A back of the daswamy, Chen Liang, Clifton McFate, diverge from Cyc’s ontology, small envelope calculation (Forbus, Liang, Matthew McLure, Irina Rabkina, and additions usually suffice to bridge the and Rabkina 2017) suggests that a low - Madeline Usher contributed heavily to gap. For example, we treat quantities as er bound of at least 2 million general - this research. fluents, as opposed to specific values, so ization pools, built up by at least 45 a single logical function suffices to million examples, would be needed to Notes 5 translate . Moreover, we have found approximate some reasonable fraction 1. In models of some processes, for exam - microtheories to be an excellent way to of what a college graduate knows. For ple, analogical matching, we capture phe - impose modularity on KB organization Companions, we will need to exploit nomena at the Cognitive band as well. and support more flexible reasoning. the data-parallel capabilities built into 2. www.cyc.com

40 AI MAGAZINE Articles

3. Here we stick to models that use the Day, S. and Gentner, D. 2007. Noninten - MAC/FAC: A Model of Similarity-Based entire Companion cognitive architecture. tional Analogical Inference in Text Com - Retrieval. Cognitive Science 19(2)(April– Other models indicate that these same ideas prehension. Memory and Cognition 35(1): June): 141–205. doi.org/10.1207/s15516709 are relevant in modeling visual analogies, 39–49. doi.org/10.3758/BF03195940 cog1902_1 including geometric analogies (Lovett et al. Dehghani, M.; Tomai, E.; Forbus, K.; and Forbus, K.; Klenk, M.; and Hinrichs, T. 2009. 2009b), oddity tasks (Lovett and Forbus Klenk, M. 2008. An Integrated Reasoning Companion Cognitive Systems: Design 2011), and Ravens’ Progressive Matricies Approach to Moral Decision-Making. In Goals and Lessons Learned So Far. IEEE (Lovett et al. 2017). Proceedings of the Twenty-Third AAAI Confer - Intelligent Systems 24(4): 36–46. doi.org/10. 4. www.freeciv.org. ence on Artificial Intelligence (AAAI). Palo 1109/ MIS.2009.71 5. The logical function QPQuantityFn we Alto, CA: AAAI Press. Forbus, K.; Liang, C.; and Rabkina, I. 2017. introduced has as its domain Cyc continu - diSessa, A.; Gillespie, N.; and Esterly, J. Representation and Computation in Cogni - ous quantities and has as its range QP theo - 2004. Coherence versus Fragmentation in tive Models . Topics in Cognitive Science. ry fluents. the Development of the Concept of Force. doi.org/10.1111/tops.12277 Cognitive Science 28: 843–900. doi.org/10. Forbus, K.; Usher, J.; Lovett, A.; Lockwood, 1207/s15516709cog2806_1 K.; and Wetzel, J. 2011. CogSketch: Sketch References Forbus, K. 1984. Qualitative Process Theory. Understanding for Cognitive Science Anderson, J. R. 2009. How Can the Human Artificial Intelligence 24(1-3): 85–168. doi. Research and for Education. Topics in Cogni - Mind Occur in the Physical Universe? Oxford, org/10.1016/0004-3702(84) 90038-9 tive Science 3(4): 648–666. doi.org/10.1111/ UK: Oxford University Press. Forbus, K. 2001. Exploring Analogy in the j.1756-8765.2011.01149.x Barbella, D., and Forbus, K. 2011. Analogi - Large. In Analogy: Perspectives from Cognitive Friedman, S. E. and Forbus, K. 2010. An cal Dialogue Acts: Supporting Learning by Science, ed. D. Gentner, K. Holyoak, and B. Integrated Systems Approach to Explana - Reading Analogies in Instructional Texts. In Kokinov. Cambridge, MA: The MIT Press. tion-Based Conceptual Change. In Proceed - Proceedings of the Twenty-Fifth AAAI Confer - Forbus, K., 2014. Qualitative Reasoning. In ings of the Twenty-Fourth AAAI Conference on ence on Artificial Intelligence (AAAI 2011). Computing Handbook, 3rd Edition: Comput - Artificial Intelligence. Palo Alto, CA: AAAI Palo Alto, CA: AAAI Press. er Science and Software Engineering. Boca Press. Barbella, D, and Forbus, K. 2013. Analogical Raton, FL: CRC Press. doi.org/10.1201/ Friedman, S., and Forbus, K. 2011. Repair - Word Sense Disambiguation. Advances in b16812-41 ing Incorrect Knowledge with Model For - Cognitive Systems 2: 297–315. Forbus, K. 2015. Analogical Abduction and mulation and Metareasoning. In Proceedings of the 22nd International Joint Conference on Barbella, D., and Forbus, K. 2015. Exploit - Prediction: Their Impact on Deception. In ing Connectivity for Case Construction in Deceptive and Counter-Deceptive Machines: Artificial Intelligence, Barcelona, Spain. Palo Learning by Reading. Advances in Cognitive Papers from the AAAI Fall Symposium. Tech - Alto, CA: AAAI Press. Systems 4: 169–186. nical Report FS-15-03. Palo Alto, CA: AAAI Friedman, S.; Barbella, D.; and Forbus, K. Press. Bello, P., and Bridewell, W. 2017. There Is 2012. Revising Domain Knowledge with No Agency Without Attention. AI Magazine Forbus, K. 2016. Software Social Organisms: Cross-Domain Analogy. Advances in Cogni - 38(4). doi.org/10.1609/aimag. v38i4.2742 Implications for Measuring AI Progress. AI tive Systems 2: 13–24. Magazine 37(1): 85–90. doi.org/10.1609/ Blass, J., and Forbus, K. D. 2015. Moral Deci - Friedman, S. E.; Forbus, K. D.; and Sherin, B. aimag.v37i1.2648 sion-Making by Analogy: Generalizations 2011. How Do the Seasons Change? Creat - Versus Exemplars. In Proceedings of the Twen - Forbus, K., and de Kleer, J. 1994. Building ing and Revising Explanations via Model ty-Ninth AAAI Conference on Artificial Intelli - Problem Solvers. Cambridge, MA: The MIT Formulation and Metareasoning. Paper pre - gence. Palo Alto, CA: AAAI Press. Press. sented at the 25th International Workshop on Qualitative Reasoning, Barcelona, Spain, Blass, J. A., and Forbus, K. D. 2016. Model - Forbus, K.; Nielsen, P.; and Faltings, B. 1991. 16–18 July. ing Commonsense Reasoning via Analogi - Qualitative Spatial Reasoning: The CLOCK cal Chaining: A Preliminary Report. In Pro - Project. Artificial Intelligence 51 (1–3): 417– Gentner, D. 1983. Structure-Mapping: A ceedings of the 38th Annual Meeting of the 471. doi.org/10.1016/0004-3702(91)90116-2 Theoretical Framework for Analogy. Cogni - Cognitive Science Society, Philadelphia, PA, Forbus, K.; Riesbeck, C.; Birnbaum, L.; Liv - tive Science 7(2): 155–170. doi.org/10.1207/ August. Austin, TX: Cognitive Science Soci - ingston, K.; Sharma, A.; and Ureel, L. 2007. s15516709cog0702_3 ety Inc. Integrating Natural Language, Knowledge Gentner, D.; Brem, S.; Ferguson, R. W.; Chang, M. D. 2016. Capturing Qualitative Representation and Reasoning, and Analog - Markman, A. B.; Levidow, B. B.; Wolff, P.; Science Knowledge with Multimodal ical Processing to Learn by Reading. In Pro - and Forbus, K. D. 1997. Analogical Reason - Instructional Analogies. Doctoral disserta - ceedings of the Twenty-Second Conference on ing and Conceptual Change: A Case Study tion, Northwestern University, Department Artificial Intelligence. Palo Alto, CA: AAAI of Johannes Kepler. The Journal of the Learn - of Electrical Engineering and Computer Sci - Press. ing Sciences 6(1): 3–40. doi.org/10.1207/ ence, Evanston, Illinois. Forbus, K.; Ferguson, R.; Lovett, A.; and s15327809jls0601_2 Cohn, A., and Renz, J. 2008. Qualitative Gentner, D. 2016. Extending SME to Han - Goldstone, R. L.; Medin, D. L.; and Gentner, Spatial Representation and Reasoning. In dle Large-Scale Cognitive Modeling. Cogni - D. 1991. Relational Similarity and the Non - Handbook of Knowledge Representation, ed. F. tive Science 41(5): 1152–1201. doi.org/10. independence of Features in Similarity Judg - van Hermelen, V. Lifschitz, B. Porter, 551– 1111/cogs. 12377 ments. Cognitive Psychology 23(2): 222–264. 596. Amsterdam, The Netherlands: Elsevier. Forbus, K. and Gentner, D. 1997. Qualita - doi.org/10.1016/0010-0285(91)90010-L doi.org/10.1016/ S1574-6526(07)03013-1 tive Mental Models: Simulations Or Memo - Hinrichs, T., and Forbus, K. 2011. Transfer Crouse, M., and Forbus, K. 2016. Elemen - ries? Paper presented at the Eleventh Inter - Learning Through Analogy in Games. AI tary School Science as a Cognitive System national Workshop on Qualitative Magazine 32(1): 72–83. doi.org/10.1609/ Domain: How Much Qualitative Reasoning Reasoning, Cortona, Italy, June 3-6. aimag.v32i1.2332 Is Required? Advances in Cognitive Systems 4. Forbus, K.; Gentner, D.; and Law, K. 1995. Hinrichs, T., and Forbus, K. 2012a. Learning

WINTER 2017 41 Articles

Qualitative Models by Demonstration. In Conference on Artificial Intelligence, New Mitchell, T.; Cohen, W.; Hruschka, E.; Taluk - Proceedings of the 26th AAAI Conference on York, NY USA. Palo Alto, CA: AAAI Press. dar, P.; Betteridge, J.; Carlson, A.; Dalvi, B.; Artificial Intelligence, 207–213. Palo Alto, CA: Lockwood, K., and Forbus, K. 2009. Multi - Gardner, M.; Kisiel, B.; Krishnamurthy, J.; AAAI Press. modal Knowledge Capture from Text and Lao, N.; Mazaitis, K.; Mohamed, T.; Nakas - Hinrichs, T., and Forbus, K. 2012b. Toward Diagrams. In Proceedings of the Fifth Interna - hole, N.; Platanios, E.; Ritter, A.; Samadi, M.; Higher-Order Qualitative Representations. tional Conference on Knowledge Capture Settles, B.; Wang, R.; Wijaya, D.; Gupta, A.; Paper presented at the Twenty-Sixth Inter - (KCAP-2009). New York: Association for Chen, X.; Saparov, A.; Greaves, M.; Welling, national Workshop on Qualitative Reason - Computing Machinery. doi.org/10.1145/ J. 2015. Never-Ending Learning. In Proceed - ing, Los Angeles CA, July 16–18. 1597735.1597747 ings of the Twenty-Ninth AAAI Conference on Artificial Intelligence. Palo Alto, CA: AAAI Hinrichs, T., and Forbus, K. 2014. X goes Lovett, A., and Forbus, K. 2017. Modeling Press. First: Teaching Simple Games Through Mul - Visual Problem Solving as Analogical Rea - timodal Interaction. Advances in Cognitive soning. Psychological Review 124(1): 60–90. Newell, A. 1990. Unified Theories of Cogni - Systems 3: 31–46. doi.org/10.1037/rev0000039 tion. Cambridge, MA: Harvard University Press. Hinrichs, T., and Forbus, K. 2016. Qualita - Lovett, A., and Forbus, K. 2011. Cultural tive Models for Strategic Planning. Advances Commonalities and Differences in Spatial Sagi, E.; Gentner, D.; and Lovett, A. 2012. in Cognitive Systems 4: 75–92. Problem-Solving: A Computational Analy - What Difference Reveals About Similarity. Cognitive Science 36(6): 1019–1050. doi.org/ Ioannides, C., and Vosniadou, S. 2002. The sis. Cognition 121(2): 281–287. doi.org/10. 10.1111/j.1551-6709.2012.01250.x Changing Meanings of Force . Cognitive Sci - 1016/j.cognition.2011.06.012 ence Quarterly 2: 5–61. Lovett, A.; Tomai, E.; Forbus, K.; and Usher, Scheutz, M. 2017. The Case for Explicit Eth - ical Agents. AI Magazine 38(4). doi.org/10. Kahneman, D. 2011. Thinking, Fast and J. 2009b. Solving Geometric Analogy Prob - 1609/aimag.v38i4.2746 s Slow. London: Allen Lane. lems Through Two-Stage Analogical Map - Tomai, E., and Forbus, K. 2007. Plenty of Kandaswamy, S.; Forbus, K.; and Gentner, ping. Cognitive Science 33(7): 1192–1231. Blame to Go Around: A Qualitative D. 2014. Modeling Learning via Progressive doi.org/10.1111/j.1551-6709.2009.01052.x Approach to Attribution of Moral Responsi - Alignment Using Interim Generalizations. Markman, A. B., and Gentner, D. 1993. bility. Paper presented at the Twenty-First In Proceedings of the 36th Annual Meeting of Structural Alignment during Similarity Workshop on Qualitative Reasoning, the Cognitive Science Society, Quebec City, Comparisons. Cognitive Psychology 25(4): Aberystwyth, UK, 27–29 June. Quebec, Canada, 23–26 July. Austin, TX: 431–467. doi.org/10.1006/cogp.1993.1011 Wilson, J.; Forbus, K.; and McLure, M. 2013. Cognitive Science Society Inc. Markman, A. B., and Gentner, D. 1996. Am I Really Scared? A Multiphase Compu - Klenk, M., and Forbus, K. 2009. Analogical Commonalities and Differences in Similarity tational Model of Emotions. In Proceedings Model Formulation for AP Physics Prob - Comparisons. Memory and Cognition 24(2): of the Second Conference on Advances in Cog - lems. Artificial Intelligence , 173(18), 1615– 235–249. doi.org/10.3758/BF03200884 nitive Systems, Baltimore, MD, 12–14 1638. doi.org/10.1016/j.artint.2009.09.003 Markman, A. B., and Medin, D. L. 2002. December. Austin, TX: Cognitive Science Klenk, M., and Forbus, K. 2013. Exploiting Decision Making. In Stevens’ Handbook of Society Inc. Persistent Mappings in Cross-Domain Ana - Experimental Psychology 3rd edition. New Winston, P. H. 1970. Learning Structural logical Learning of Physical Domains. Arti - York: John Wiley & Sons, Inc. doi.org/ Descriptions From Examples. Ph.D. diss., ficial Intelligence 195: 398–417. doi.org/10. 10.1002/0471214426.pas0210 Computer Science, Massachusetts Institute 1016/j.artint.2012.11.002 McFate, C., and Forbus, K. 2016a. Analogi - of Technology, Cambridge, MA. Kuehne, S.; Gentner, D.; and Forbus, K. cal Generalization and Retrieval for Denom - 2000. Modeling Infant Learning via Sym - inal Verb Interpretation. In Proceedings of the bolic Structural Alignment. In Proceedings of 38th Annual Meeting of the Cognitive Science Kenneth D. Forbus is the Walter P. Murphy the 22nd Annual Meeting of the Cognitive Sci - Society, Philadelphia, PA, 10–13 August. Professor of Computer Science and a pro - ence Society, Philadelphia, PA, August. Austin, TX: Cognitive Science Society, Inc. fessor of education at Northwestern Univer - Austin, TX: Cognitive Science Society Inc. McFate, C., and Forbus, K. 2016b. An Analy - sity. His research interests include qualita - Laird, J. 2012. The SOAR Cognitive Architec - sis of Frame Semantics of Continuous tive reasoning, analogical reasoning and ture. Cambridge, MA: The MIT Press Processes. In Proceedings of the 38th Annual learning, natural langauge understanding, Laird, J.; Lebiere, C.; and Rosenbloom, P. Meeting of the Cognitive Science Society, sketch understanding, and cognitive archi - 2017. A Standard Model of the Mind: Philadelphia, PA, 10–13 August. Austin, TX: tecture. Cognitive Science Society, Inc. Toward a Common Computational Frame - Thomas Hinrichs is a research associate McFate, C. J.; Forbus, K.; and Hinrichs, T. work Across Artificial Intelligence, Cogni - professor of computer science at North - 2014. Using Narrative Function to Extract tive Science, Neuroscience, and Robotics. AI western University. His research interests Qualitative Information from Natural Lan - Magazine 38(4). doi.org/10.1609/aimag. include commonsense reasoning, cognitive guage Texts. In Proceedings of the Twenty- v38i4.2744 architectures and machine learning. Liang, C., and Forbus, K. 2015. Learning Eighth AAAI Conference on Artificial Intelli - Plausible Inferences from Semantic Web gence. Palo Alto, CA: AAAI Press. Knowledge by Combining Analogical Gen - McLure, M. D.; Friedman S. E.; and Forbus, eralization with Structured Logistic Regres - K. D. 2015. Extending Analogical General - sion. In Proceedings of the Twenty-Ninth AAAI ization with Near-Misses. In Proceedings of Conference on Artificial Intelligence. Palo Alto, the Twenty-Ninth AAAI Conference on Artifi - CA: AAAI Press. cial Intelligence. Palo Alto, CA: AAAI Press. Liang, C., Paritosh, P., Rajendran, V., and McShane, M. 2017. Natural Language Forbus, K. D. 2016. Learning Paraphrase Understanding (NLU, not NLP) in Cogni - Identification with Structural Alignment. In tive Systems. AI Magazine 38(4). doi.org/ Proceedings of the 25nd International Joint 10.1609/aimag.v38i4.2745

42 AI MAGAZINE