A Manifesto for an Integrative Design-oriented Approach to Understanding Humans as Autonomous Agents

This is a draft manifesto for an integrative design-oriented (IDO) approach to understanding humans as autonomous agents (“cognitive science”, “AI”, “psychology”, etc.)

Researchers are invited to contribute/participate.

This manifesto (in progress) is meant to be a collaborative effort for publication in a journal like Artificial General Intelligence (though it applies primarily to understanding natural autonomous agency through theoretical computational paradigms). It is meant to influence how at least some research is planned, funded and conducted. It would hopefully spur new grant programmes and criteria for existing grant programmes — for governments, commercial research and private research foundations.

This manifesto might migrate to its own domain, at which point this web page will link to it.

(Compare another co-authored manifesto of which Beaudoin is the editor: Manifesto for Ubiquitous Linking.)

Context

This manifesto may be read as a response to the following problems in cognitive science, AI and psychology.

IMPORTANT NOTE: The following are some of the limitations of current research. Limitations are inevitable. This list of limitations is not meant to be read as personal criticisms of researchers in the field. Nor is it meant to suggest that prior R&D was deeply flawed. It is instead meant to the reality to which the IDO is a response. IDO is not meant to be “the holy grail” or the only approach to use. An IDO approach needs to benefit from the the products of the research approaches that it complements.

  1. Cognitive science is supposedly the interdisciplinary study of all aspects of the human mind using computational metaphors. However, rarely are affective, conative/motivational, and ancillary functions (such as sleep) considered under the banner of cognitive science. For instance, at Cognitive Science conferences there are sometimes symposia on emotion but rarely is affect considered in detail outside such symposia. And even when affect is discussed, it tends to be treated as something separate from cognition that influences or may be influenced by cognition. Rarely are cognition and affective processes considered as blended , let alone with motivational and ancillary ones.
  2. The expression cognitive science is potentially misleading because although cognitive science is described broadly to include all natural information-processing, it tends to be identified with cognition as dry and separate from other types, aspects or functions of information processing.
  3. “Affective science” has emerged. One might expect motivation in affective science circles to be considered of great interest, yet it tends to be neglected. It is treated at best as a separate “component” of emotion, or something that is influenced by affect. Theories of affect , mood and emotion don’t tend to explicitly mention particular theories of motivation.
  4. Scientific and engineering research & development have increasingly become specialized and narrow in focus. It is admittedly very difficult to consider the “big picture” and deal with hard, integrative problems while making measurable progress. Requirements to publish, get one’s work funded, achieve tenure and further promotion, and build one’s reputation all get in the way of integrative efforts; as do administrative loads, teaching requirements and family responsibilities. Moreover, science has fads and zealots which are not aligned with IDO (e.g., embodied cognition and associationism).
  5. Theoretical cognitive scientists, still relatively few in number, tend to work individually or in small groups. They almost never employ a professional software architect. In contrast, development of large commercial software involves dozens or hundreds of developers and a team of software architects. Yet commercial software is infinitely simple compared to the mind/brain. Software architects in commercial projects strive to ensure that the software product meets the overall requirements of stakeholders and is intelligible; that trade-offs are well-understood, and so forth. Such considerations have analogs in the development of computational theories of autonomous agency.
  6. As promising, important, successful (with certain classes of problems, and blindingly popular as associative/statistical AI (deep learning, connectionism, etc.) may be, it is not a sufficient framework for addressing the hard problems of theoretical and applied AI. For example, much of human competence is in essential ways innate, and not practically computable via statistical AI. Much competence needs to be programmed into theory and models as priors that are merely parameterized by the environment. For example, the perception of impossibility and necessity cannot fully be explained by statistical models — nor (incidentally) can they all be explained by derivation from definitions and Fregean logic (also a frequent foil of connectionism): some synthetic knowledge is known / provable without logical deduction from arbitrary definition and without being an empirical generalization. For alternatives to purely statistical AI and formal logic see the meta-morphogenesis project, Rebooting AI by Gary Marcus & Ernest Davis, and the work of Judea Pearl; there are others.
  7. Whether there is a replication crisis in psychology or not, psychology has been held back by (a) lack of concern for integrative design-oriented theory that can explain competence and (b) an extreme willingness and tendency to run empirical studies in the absence of concern for such theory. Theoretical terms and psychometric instruments are often used willy nilly. (See A problem in theory; Newell’s “You can’t play 20 questions”; Beaudoin et al 2017, Ekkekakis (2013)The Measurement of Affect, Mood, and Emotion; etc.)
  8. While research on “cognitive architectures” is promising, much of it to date tends unfortunately to be true to the ancient definition of the first part of its name (cognitive), ignoring affective, motivational and ancillary requirements. Also it tends to overlook the fact that in some species (e.g., humans) information processing architectures (normally) develop (are not static). Positing architectures, while not merely helpful but necessary for modeling autonomous agency, can lead to problematic simplifications and rigid assumptions. See Pessoa’s arguments in The Cognitive Emotional Brain for embracing complexity; and Sloman’s arguments to recognize multiple discontinuities (rather than simplistic sharp dichotomies) .

Fundamentals of IDO (Integrative design-oriented approach to cognitive science and AI).

To be clarified and slightly expanded.

The fundamentals of an IDO approach are as follows.

  1. The IDO approach is ultimately concerned with explaining and replicating autonomous agency (Beaudoin, 1994).
  2. The IDO approach seeks to explains competence. It is more concerned with what (actual or possible) systems can do than with predicting what they will actually do in particular circumstances.

  3. The IDO approach seeks to understand mental phenomena from a design stance (or designer stance). Specifying environmental niches and requirements for systems; specifying designs to meet those requirements, implementing designs; analyzing and evaluating the extent to which the designs meet the requirements, and that the implementations satisfy the designs and their requirements; repeating this process.

  4. This means the IDO approach seeks to produce information-processing theories.
  5. The IDO approach while accepting the relevance of empirical data and phenomena-based methods, rejects both inductivism and unbridled empiricism. (Compare Muthukrishna & Henrich, 2019: A problem in theory).
  6. The IDO approach can and should be used not merely to directly understand autonomous agency directly, but to study any and all types of specific human capabilities and mechanisms. For instance, this approach can and should be used to understand the human sleep onset control system, or vision.
  7. Thus, IDO research often seeks to understand specific, arguably narrow, functionality (e.g., an aspect of vision, the sleep-onset control system, decision-making); but in so doing it explicitly draws on larger IDO theories/research. For example, an IDO theory of the human sleep-onset control system might draw on integrative theories of consciousness and motive processing. More generally, when using theoretical terms (e.g., “arousal” or “attention”), IDO research attempts to ground such terms in theories, preferably IDO theories, which make the terms more meaningfully interpretable, as opposed to using the terms and leaving it up to the reader to guess what the IDO meaning of the term might be. For example, one might reference Donald (2001) to ground one’s concept of consciousness. This means that IDO theorizing involves at least provisional commitment to theory.
  8. The IDO approach ultimately seeks to explain the integration (interaction and even blending) of major information processing functions: cognitive, motivational, affective, executive, ancillary, etc.
  9. The IDO approach is sincerely interdisciplinary. Key members of IDO R&D projects are well versed (and publish in) several of the core IDO disciplines, such as AI, philosophy, psychology, neuroscience, linguistics and anthropology. This ideally is augmented with multidisciplinary collaboration (involving experts representing a variety of fields).
  10. The IDO approach recognizes the need to explore, develop, implement and assess information processing architectures. The study of information processing architectures is itself a demanding discipline. Therefore IDO research teams preferably include or designate AI/software/system architects, whether part of the core team or as consultants, advisors and/or reviewers.
  11. The IDO approach recognizes that there is not only a single (information processing) architecture that supports all forms of intelligence. There is a space of possible minds. For instance, some species support multiple information processing architectures; and in some species some individuals’ information processing architecture develops over time.
  12. The IDO approach is comparative: It analyzes the space of possible niches, requirements, designs, and implementations, relations between them, and tradeoffs.
  13. The IDO approach recognizes Artificial General Intelligence (AGI) as a general science of intelligence / competence. Compare Prospects paper. AI provides, or should provide, theoretical and methodological frameworks for modeling the space of possible minds, including those of all present, past, future and possible actual minds.
  14. The IDO approach while focusing on information processing is multi-scale, bridging genes to competence. Compare Grant, 2003: “Systems biology in neuroscience: bridging genes to cognition”). The TCP/IP stack and virtual machines in commercial and open-source software illustrate an extremely wide variety of possibilities of information-processing layering. Virtual machine concepts and causality are relevant to understanding autonomous agents. It is at least worth exploring the (often neglected possibility) that human brains host layered, interconnected virtual machines.
  15. The IDO approach recognizes the need to understand the evolution of information-processing and competence. (For example the meta-morphogenesis project, as an offshoot of Alan Turing’s final paper. See 2013: A problem in theory) for an argument for evolutionary grounding.
  16. The IDO approach recognizes that, as mentioned in the preamble, associative mechanisms are not sufficient for modeling sophisticated autonomous agency, though they may be necessary. “Kantian” mechanisms and as yet unexplored representations and mechanisms will be required.
  17. As key examples of its commitment to one of the core disciplines, philosophy: The IDO approach recognizes (a) the need for conceptual analysis for understanding mental phenomena and clarifying concepts and augmenting/contextualizing factor analysis; (b) the importance of thought experiments.

  18. While the IDO approach is not committed to any specific theory of mind/brain, it admits the possible integrative, mechanistic relevance of concepts of consciousness (E.g.,. Donald, M. (2001). A mind so rare: The evolution of human consciousness).

  19. IDO researchers acknowledge the IDO shortcomings of their own scientific communications. This means they recognize ways in which their own theories fail to meet the IDO objectives. (Humility)

Contextual quotes

The opening quotations of my 1994 Ph.D. thesis are still relevant

The problem is not that we do not know which theory is correct, but rather that we cannot construct any theory at all which explains the basic facts (Power, 1979 p. 109)

I think that when we are speculating about very complicated adaptive systems, such as the human brain and social systems, we should especially beware of oversimplification—I call such oversimplification “Ockham’s lobotomy”. (Good, 1971a p. 375)

Some relevant literature

One of the first papers to use the expression integrative design-oriented is Beaudoin et al (2022) Mental perturbance: An integrative design-oriented concept for understanding repetitive thought, emotions and related phenomena involving a loss of control of executive functions .

NB: The IDO approach was described in Cognitive Productivity: Using Knowledge to Become Profoundly Effective. However, there Beaudoin used the term “design-based”, following Aaron Sloman and Beaudoin’s terminology of the early 1990s. However, “design-based” developed a very different meaning in psychology. “Design-oriented” is preferable.

Authorship and revisions

This is intended to lead to a jointly authored document.

  • 2022-10-26. Updated based on some of the feedback by Alice Dauphin and Guillaume Pourcel
  • 2021-07-16. Merged two bibliographical sections, minor edits.
  • 2021-01-13. Title and reference and minor changes.
  • 2019-09-22. An earlier version of this manifesto written by Luc Beaudoin was presented at the 2019 World Sleep Congress in Vancouver, BC.
  • 2019-11-17. Luc P. Beaudoin. First draft of the manifesto post 2019-World Sleep Congress.
  • 2019-11-18. Luc P. Beaudoin. Added paragraphs to the pre-amble and to the manifesto (regarding evolution and multi-scales), and a reference.

A blog post introducing this manifesto..

Feedback and signatures

Please send feedback to lucb@cogzest.com, use the comments section below, or use Twitter.

Acknowledgements

Thanks to Alice Dauphin and Guillaume Pourcel for feedback on this document. 2021-07-16: Some of their feedback still needs to be incorporated in this document

2 thoughts on “A Manifesto for an Integrative Design-oriented Approach to Understanding Humans as Autonomous Agents”

  1. #1. General comments…

    We made those comments as if you intend to write the manifesto for anyone doing AGI or cognitive science (including psychologists). In particular, we imagined that anyone of those disciplines should be able to understand it and potentially to change its practice to align with IDO.

    ##1. …about the content 

    – As we are familiar with IDO and Sloman taxonomy, we had no problem to understand your points but it is less sure that a naive audience would understand you in details. 
    – Maybe you should add a clear definition of “IDO” and “design stance” somewhere.
    – Give more examples to illustrate and convince people. In our case it was thanks to the examples that we were convinced and that we could understand the potential of the design stance (such as the rich explicit/implicit distinction that comes from AI/computer science).

    – From what we understand, your purpose is the following: (1) to widen the subjects studied by cognitive science (in particular to include emotions, motivations… to the “dry” aspects of cognition), (2) that cognitive scientists adopt the design stance and consequently understand and admit that their research is part of a broader research aiming at understanding the space of possible minds.
    – Maybe these 2 points (if they are indeed your main points) should appear clearly (as an introduction ?) (a) to clarify the document and (b) to help the reader to link each argument to the corresponding point. 
    – Why do you expect people to adopt the design stance is not explicitly mentioned and we think it should be detailed for naive readers as it is the core point of the manifesto.
    – Design stance is a mean to define new relevant concepts and concepts are key issues in cognitive science.
    – Using the design stance, some pitfalls can be avoided such as overfitting and missing some aspects of competences. 
    – Adopting the design stance is acknowledging the difficulty of the endeavour, split it and make space for theory. For us it’s one of the key insights of Minksy and Sloman that changed our way to see cognitive science: we need to give space to the theory to avoid simplifying views (Ockham lobotomy).
    – With the design stance, the mechanisms of intelligence do not compete with each other but inhabit the space of agents capacities. Consequently, some debates (such as associationists methods VS non-associationists) become obsolete.
    – We also think that more emphasis should be done in evaluating the trade-off of the models presented. A lot of work consists in developing some frameworks (Free-energy principle, Reinforcement Learning…) while not focusing on the qualitative differences in terms of competences. In general benchmarking is a key challenge for cognitive computational science.

    ##1.2 …about the structure

    – To help naive readers, it seems to us that few explicit proposals on which you elaborate with examples could be better understood than a list. Or maybe the list should be organised to bring out the main points. For example, the points 1, 2, 3 in the preamble could appear under the same umbrella “Cognitive science should go back to its initial goal, that is explain all aspects of cognition, not only the dry aspects of cognition”.
    – We also think that the manifesto could emphasize that a lot of the IDO approach can be done quite easily without requiring a lot of investment, just by seeing the problem from the design stance.


    ##1.3 …about the tone

    – As the purpose is to change people’s vision of their field, we think that the tone should be encouraging.
    – For that, maybe write the manifesto as the first step toward an ideal for cognitive science can be less pejorative and people will feel less pointed out. 
    – Some points you suggest to change are (we think) acknowledged by researchers but still difficult to implement:
    – Software developer: difficult because of the lack of funding
    – True interdisciplinarity: difficult because of the vast amount of knowledge and methodologies to master. Maybe the point might be to encourage truly interdisciplinary collaboration.

    #2 Comments addressing specific points

    ##2.1 Preamble

    – (6.) *“Theoretical cognitive scientists and AI researchers, still relatively few in number, tend to work individually or in small groups”*
    – This point made us think to a more general problem existing in research, the fact that a lab is made of a PI, his postdoc and phD students (maybe this is a scheme specific to France, we are not really sure for other countries). This is problematic as to elaborate on a complex problem it is necessary to have people expert in different domains. This is why people do collaborations with other labs but these collaborations are constrained by spatio temporal limits. To have labs with several mature researchers and engineers interacting seems more promising.
    – (9.) *“While research on cognitive architectures is promising, much of it to date tends to ignore affective and motivational requirements.”* 
    – We agree, but we see a significant trend toward integrating motivation in cognitive architectures with the work of J. Bach on MicroPsi, R. Sun on Clarion, eBICA from the BICA community led by A. Samsonovich…


    ##2.2 Fundamentals of IDO

    – (1.) The name Artificial General Intelligence (AGI) is quite blurry. Especially because it is also a specific community with their own vision.
    – (4.) Maybe a detailed explanation of the limitations of a factor analysis could help understand the relevance of the IDO approach.
    – (5.) Same as 4.
    – (8.) *“The IDO approach is truly interdisciplinary”* 
    – Maybe to explain the consequences of IDO approach on interdisciplinary can help. For example, IDO can provide coherent definitions across domains, allowing researchers to speak the same (computational) language. It might also be interesting to point the advantage of it.

Leave a Reply

This site uses Akismet to reduce spam. Learn how your comment data is processed.