Selmer Bringsjord’s research while affiliated with Rensselaer Polytechnic Institute and other places

The so-called symbol-grounding problem (SGP) has long plagued cognitive robotics (and AI). If Rob, a humanoid household robot, is asked to remove and discard the faded rose from among the dozen in the vase, and accedes, does Rob grasp the formulae/data he processed to get the job done? Does he for instance really understand the formulae inside him that logicizes “There’s exactly one faded rose in the vase”? Some (e.g., Searle, Harnad, Bringsjord) have presented and pressed a negative answer, and have held that engineering a robot for whom the answer is ‘Yes’ is, or at least may well be, insoluble. This negativity increases if Rob must understand that giving a faded rose to someone as a sign of love might not be socially adept. We change the landscape, by bringing to bear, in a cognitive robot, an unprecedented, intertwined quartet of capacities that make all the difference: namely, (i) social planning; (ii) multi-modal perception; (iii) pre-meditated attention to guide such perception; and (iv) automated defeasible reasoning about causation. In other words, a genuinely social robot that senses in varied ways under the guidance of how it directs its attention, and adjudicates among competing arguments for what it perceives, solves SGP, or at least a version thereof. An exemplar of such a robot is our PERI.2, which we demonstrate in an environment called ‘Logi-Forms,’ intelligent navigation of which requires social reasoning.

Artificial intelligence and machine learning are increasingly used to offload decision making from people. In the past, one of the rationales for this replacement was that machines, unlike people, can be fair and unbiased. Evidence suggests otherwise. We begin by entertaining the ideas that algorithms can replace people and that algorithms cannot be biased. Taken as axioms, these statements quickly lead to absurdity. Spurred on by this result, we investigate the slogans more closely and identify equivocation surrounding the word 'bias.' We diagnose three forms of outrage-intellectual, moral, and political-that are at play when people react emotionally to algorithmic bias. Then we suggest three practical approaches to addressing bias that the AI community could take, which include clarifying the language around bias, developing new auditing methods for intelligent systems, and building certain capabilities into these systems. We conclude by offering a moral regarding the conversations about algorithmic bias that may transfer to other areas of artificial intelligence.

We propose an extension to Legg and Hutter’s universal intelligence (UI) measure to capture the intelligence of agents that operate in uncomputable environments that can be classified on the Arithmetical Hierarchy. Our measure is based on computable environments relativized to a (potentially uncomputable) oracle. We motivate our metric as a natural extension to UI that expands the class of environments evaluated with a trade-off of further uncomputability. Our metric is able to capture intelligence of agents in uncomputable environments we care about, such as first-order theorem proving, and also lends itself to providing a notion of intelligence of oracles. We end by proving some properties of the new measure, such as convergence (given certain assumptions about the complexity of uncomputable environments).

We lay some groundwork for a family of theorems that, in general, assert that as some factors internal to an artificial agent increase or decrease on their respective scales, where these factors seem to contribute to an increase in autonomy overall, the level of trust on the part of a human agent that this artificial agent will in fact deliver on this trust declines.

To diagnose someone’s reasoning today as “sophistry” is to say that this reasoning is at once persuasive (at least to a significant degree) and logically invalid. We begin by explaining that, despite some recent scholarly arguments to the contrary, the understanding of ‘sophistry’ and ‘sophistic’ underlying such a lay diagnosis is in fact firmly in line with the hallmarks of reasoning proffered by the ancient sophists themselves. Next, we supply a rigorous but readable definition of what constitutes sophistic reasoning (=sophistry). We then discuss “artificial” sophistry: the articulation of sophistic reasoning facilitated by artificial intelligence (AI) and promulgated in our increasingly digital world. Next, we present, economically, a particular kind of artificial sophistry, one embodied by an artificial agent: the lying machine. Afterward, we respond to some anticipated objections. We end with a few speculative thoughts about the limits (or lack thereof) of artificial sophistry, and what may be a rather dark future.

Research in automated planning traditionally focuses on model-based approaches that often sacrifice expressivity for computational efficiency. For artificial agents that operate in complex environments, however, frequently the agent needs to reason about the beliefs of other agents and be capable of handling uncertainty. We present Spectra, a STRIPS-inspired AI planner built atop automated reasoning. Our system is expressive, in that we allow for state spaces to be defined as arbitrary formulae. Spectra is also designed to be logic-agnostic, as long as an automated reasoner exists that can perform entailment and question-answering over it. Spectra can handle environments of unbounded uncertainty; and with certain non-classical logics, our system can create plans under epistemic beliefs. We highlight all of these features using the cognitive calculus $\mathcal {DCC}$. Lastly, we discuss that under this framework, in order to fully plan under uncertainty, a defeasible (= non-monotonic) logic can be used in conjunction with our planner.

Formal deductive logic, used to express and reason over declarative, axiomatizable content, captures, we now know, essentially all of what is known in mathematics and physics, and captures as well the details of the proofs by which such knowledge has been secured. This is certainly impressive, but deductive logic alone cannot enable rational adjudication of arguments that are at variance (however much additional information is added). After affirming a fundamental directive, according to which argumentation should be the basis for human-centric AI, we introduce and employ both a deductive and—crucially—an inductive cognitive calculus. The former cognitive calculus, DCEC, is the deductive one and is used with our automated deductive reasoner ShadowProver; the latter, IDCEC, is inductive, is used with the automated inductive reasoner ShadowAdjudicator, and is based on human-used concepts of likelihood (and in some dialects of IDCEC, probability). We explain that ShadowAdjudicator centers around the concept of competing and nuanced arguments adjudicated non-monotonically through time. We make things clearer and more concrete by way of three case studies, in which our two automated reasoners are employed. Case Study 1 involves the famous Monty Hall Problem. Case Study 2 makes vivid the efficacy of our calculi and automated reasoners in simulations that involve a cognitive robot (PERI.2). In Case Study 3, as we explain, the simulation employs the cognitive architecture ARCADIA, which is designed to computationally model human-level cognition in ways that take perception and attention seriously. We also discuss a type of argument rarely analyzed in logic-based AI; arguments intended to persuade by leveraging human deficiencies. We end by sharing thoughts about the future of research and associated engineering of the type that we have displayed.

We introduce rudiments of the cognitive meta-architecture M (majuscule of $\mu$ and pronounced accordingly), and of a formal procedure for determining, with M as touchstone, whether a given cognitive architecture $X_i$ (from among a finite list 1 $\ldots k$ of modern contenders) conforms to a minimal standard model of a human-level AGI mind. The procedure, which for ease of exposition and economy in this short paper is restricted to arithmetic cognition, requires of a candidate $X_i$, (1), a true biconditional expressing that for any human-level agent a, a property possessed by this agent, as expressed in a declarative mathematical sentence s(a), holds if and only if a formula $\chi _i(\mathfrak {a})$ in the formal machinery/languages of $X_i$ holds as well ($\mathfrak {a}$ being an in-this-machinery counterpart to natural-language name a). Given then that M is such that $s(a) \text { iff } \mu (\mathfrak {m})$, where the latter formula is in the formal language of M, with $\mathfrak {m}$ the agent modeled in M, a minimal standard modeling of an AGI-level mind is certifiably achieved by $X_i$ if, (2), it can be proved that $\chi _i(\mathfrak {a}) \text { iff } \mu (\mathfrak {a}).$ We conjecture herein that such confirmatory theorems can be proved with respect to both cognitive architectures NARS and SNePS, and have other cognitive architectures in our sights.Keywordsstandard modeling of AGI-level mindscognitive architecturescomputational logic