Dynamically Adjusted Control System Architecture

chevron-icon
RFP Proposals
Top
chevron-icon
project-presentation-img
Expert Rating 3.7
Luke Mahoney (MLabs)
Project Owner

Dynamically Adjusted Control System Architecture

Expert Rating

3.7

Overview

We propose to develop a flexible motivational framework within which AGI systems can be supported. Our goal is to allow for concurrent information processing, and dynamically adjusted control, both responsive to changes in the external environment, and the internal state of the model. We are basing our framework on the Information Fusion work of Frankel and Bedworth which is detailed in the 2000 paper. We will describe our novel AGI framework in a research paper, and illustrate it in operation with a simple “Hello, world” demonstrator.

RFP Guidelines

Develop a framework for AGI motivation systems

Complete & Awarded
  • Type SingularityNET RFP
  • Total RFP Funding $40,000 USD
  • Proposals 12
  • Awarded Projects 2
author-img
SingularityNET
Aug. 13, 2024

Develop a modular and extensible framework for integrating various motivational systems into AGI architectures, supporting both human-like and alien digital intelligences. This could be done as a highly detailed and precise specification, or as a relatively simple software prototype with suggestions for generalization and extension.

Proposal Description

Company Name (if applicable)

MLabs LTD

Project details

Details

Narrow AI focuses on the forward, local process of estimation and analysis. Goals and requirements are dictated by the designer, and system accuracy is measured using metrics specified outside the system. In their seminal paper Control, Estimation and Abstraction in Fusion Architecture: Lessons from Human Information Processing the authors explicitly define the feedback, global process of control and emotion which seems to underpin the self-motivated information processing undertaken by people. We note that the authors use the anthropomorphically-laden word “emotion” in their paper - we much prefer the word “motivation” for what we believe is the same control mechanism.

They go on to demonstrate that goal setting and measurements of efficacy can be generated within such an architecture, and that such an architecture could be applied to automated fusion systems thereby bestowing it with an autonomy of reasoning, and agency in its own dynamic goal-setting and efficacy-measuring capabilities.

Our team sees this general approach as key to the widening of narrow AI applications to AGI deployments. While the authors of the original paper place the architecture in terms of use of, and control of, information assets (there is a clear hint at military intelligence activities in the paper), we see the same infrastructure as being easily adapted to a more benign environment, where the same global control architecture can be considered as an adaptive control framework for AGI, supporting the systems adaptability to, and alignment with, the environment in which it operates.

The architecture as it was first described nearly a quarter of a century ago, is already modular in not only the estimation and control loops; but also in the layering of abstraction, which is fundamental to the framework. This layered abstraction allows vertical motivation of prerequisite supply (up) and requirements / needs (down) - much like Maslow’s hierarchy. This modular form, and the associated message-passing data-flow, offer scalability for processing in very large AGI systems, when implemented on decentralized infrastructure.

Our development of the framework will build on the existing capability to dynamically prioritize motivation as its environment and its internal model of the that environment evolves, thereby enabling adaptive complex behavior.

The layered abstraction model allows for top down requirements to be defined both in response to the sensed environment, but also by an external controller. The scenario alluded to in the 2000 paper is one of command and control, but we can repurpose this ability to define external constraints to allow ethical guidelines to be integrated as one of the abstraction layers, giving the system to remain aligned with a pre-defined moral compass.

Our proposed project falls into two main parts:

  • define a novel, modular architecture for processing and motivation within AGI systems; using the framework from the 2000 paper. The new architecture will specifically address learning, and social conformity as parts of the control loop. While we are keen to build on insights from human intelligence, we do not wish to predispose the architecture to be constrained by this. Rather we wish to develop a framework which supports human intelligence as a special case, but admits other possibilities, such as alien life-based or digital intelligences
  • implement a simple prototype of this framework operating at a minimal size, to illustrate each of the components i.e. estimation and control, at two levels of abstraction. The demonstrator will be based on a very basic scenario such as learning and reasoning in a limited environment

Our outcomes will follow the same structure. We will produce a research paper describing our novel architecture in detail, and the benefits it affords to the progress of AGI. The paper will be delivered as a milestone to SingularityNET; we will be happy to present it internally to SingularityNET members, and would welcome discussion and collaboration.

Our intention is to provide a foundation for further research: we do not think we will have solved all aspects of the proposed framework within the constraints of this initial project. If there is sufficient interest we would intend to submit an updated version of this research paper to a recognized journal or conference, with authorship reflecting the efforts of our team, and any additional contributions made following collaboration.

We would also deliver a demonstrator, written in MeTTA to illustrate the application of our framework. While the framework itself is agnostic to the inference paradigms which implement each of the modules, we intend to build this initial demonstrator using Probabilistic Logic Networks for the higher level symbolic reasoning, and sub-symbolic processing for the sensor interface.

Challenges

This is an exploratory project, aiming at developing and demonstrating a novel framework for AGI, and hopes to illustrate a broad applicability to a wide variety of potential deployments. As such it is an ambitious remit for such a short project with limited budget. We believe that this risk is at least somewhat mitigated by the extensive experience of the project team members. Furthermore, we invite discussion and collaboration from the wider AGI community, and trust that this will further reduce technical risk.

Team

MLabs AI sees AGI as a very important strategic goal, and our senior technical team already devote a portion of their time in pursuit of this outcome. We see our collaborations with SingularityNET as an important part of this endeavor.

Mark Florisson, CEO and AGI Projects Champion - Project Leader

Mark Florisson is co-founder and CEO of MLabs AI, and has a strong background in theoretical computer science, and compilers. Mark also heads up the wider MLabs group as CEO, and provides project management and technical leadership for our key activities. He has a longstanding interest in AGI, using both symbolic, and sub-symbolic information processing. He is currently investigating the high-level architecture for AGI, part of which comprises this project, as well as a low-level architecture which supports purely Neural approaches to AGI. Mark will lead the project, and direct the implementation of the demonstrator..

Mark Bedworth PhD FSS, Chief Scientist and AI Visionary - Project Architect

Dr Bedworth has over 40 years experience at the forefront of AI, and is one of the authors of the 2000 AGI framework paper with US psychologist Carl Frankel. His early work included development of Boltzmann machines alongside Nobel Laureate Geoff Hinton, and the development of several novel forms of Neural Network. His work on Bayesian probability moderation mitigated the problems inherent in the veto effect, for which he received worldwide recognition in the information fusion community. He co-founded the International Society for Information Fusion, and served as its Vice President in its second year. Two of his patents have been acquired by Apple, and form a core element of the Siri speech recognition system. He is currently co-founder and Chief Scientist at MLabs AI, leading the team developing radically new approaches to Deep Learning, self-directed knowledge acquisition and Artificial General Intelligence.

Mike Moore PhD, AI Team Leader and Philosophy & Ethics Expert - Project Consultant and Software Architect

Dr Moore earned his Doctorate in philosophy from Essex University in the UK, where he studied the role of goal-setting in ethics. A self-taught software architect, and experienced machine learning practitioner, he is fascinated with the intersection of ethics and AI - a factor he believes should be central to developments in AGI. He currently leads the computer vision team at MLabs AI, as well as pioneering approaches for ensuring that AI and AGI make a positive contribution to society. Mike will guide the ethical elements of the project, and contribute to the implementation of the demonstrator.

 

Open Source Licensing

MIT - Massachusetts Institute of Technology License

Links and references

References

https://drive.google.com/file/d/1OWrk9XEL-PevOFTZUHHMA6WWjYqK2Ssm/view?usp=sharing

Website

https://www.mlabs.city/

Proposal Video

Not Avaliable Yet

Check back later during the Feedback & Selection period for the RFP that is proposal is applied to.

  • Total Milestones

    4

  • Total Budget

    $30,000 USD

  • Last Updated

    6 Dec 2024

Milestone 1 - Description of AGI Framework

Description

In this milestone we will develop the broad definitions of the AGI framework. Basing our architecture on the Frankel-Bedworth model for information fusion we will define the modular components of a forward-backward / processing-motivation AGI system operating at multiple levels of abstraction.

Deliverables

Project white paper

Budget

$10,000 USD

Success Criterion

White paper adequately describes the background to our work, the essence of the control and estimation architecture from 2000, and how we have used it as a springboard to develop a novel AGI framework.

Milestone 2 - Specification of AGI Framework

Description

In this milestone we will extend the model to cover the two key aspects of AGI: self-directed learning and socio-ethical conformity. Learning should be modularised in such a way that supports continuous improvement while maintaining the ability to generalise. Conformity to expectations from external influences will form a key element of the top-down control signal. In this milestone we will also introduce the multiple levels of abstraction with which the AGI system can operate.

Deliverables

Specification document

Budget

$5,000 USD

Success Criterion

Specification document describes the leaning and conformity elements of our novel AGI framework, together with the levels of abstraction which enable top-down motivation in addition to self-directed goal setting. We will also list a number of use-cases, and how we see them fitting into our framework. We will include ChatBots, Humanoid Robots and Virtual Agents among our use cases.

Milestone 3 - MeTTa Demo of AGI Framework

Description

For this milestone we will develop an implementation of a very basic AI system which illustrates the behavior of designs based on our framework. Detailed specifications will be finalised nearer the time but we will construct a system which encompasses sensing perception decision making and learning of the forward loop; plus expectation preference and motivation of the backward loop. We suggest a very simple 2D environment consisting of colored polygonal shapes which interact in a specified way and under the constraints of a prescribed rule system. The system will interact autonomously with its environment using reinforcement learning; and with a “teacher” using supervised learning. The idea here is not to illuminate the boundaries of what the architecture can encompass but to provide a “Hello world” illustration of how it might fit into the Hyperon AGI Framework. Key elements of even a simple demonstrator are motivation and prioritization both of which we intend to illustrate in some basic form.

Deliverables

Framework demo implemented in MeTTa

Budget

$10,000 USD

Success Criterion

Demonstrator software is functional, written in MeTTa, and adequately showcases our novel AGI framework using inferencing paradigms as necessary e.g. PLNs or NNs.

Milestone 4 - AGI Framework Research Paper

Description

This milestone comprises the production of a final version of the research paper describing the Florisson-Frankel-Bedworth AGI architecture. It will outline our findings and detail the behavior of a software implementation of the demonstrator system. The paper will be based on the white paper reports of milestones 1 and 2. It will include relevant sections on background motivation insights and experiments. We will conclude with an honest description of any limitations and remaining challenges our thoughts on how the framework might be further developed or extended and an indicative roadmap of how these might be addressed in future work.

Deliverables

Research paper detailing the Florisson-Frankel-Bedworth AGI architecture

Budget

$5,000 USD

Success Criterion

A research paper of sufficient quality is submitted, or in sufficient condition to be submitted, to a relevant AGI journal or conference. A presentation of our findings to the SingularityNET community also organised and available if requested.

Join the Discussion (0)

Expert Ratings

Reviews & Ratings

Group Expert Rating (Final)

Overall

3.7

  • Feasibility 3.5
  • Desirabilty 4.0
  • Usefulness 4.0
  • Expert Review 1

    Overall

    1.0

    • Compliance with RFP requirements 1.0
    • Solution details and team expertise 1.0
    • Value for money 1.0
    Information Fusion by Frankel and Bedworth

    This project is fully based on the work regarding Information Fusion by Frankel and Bedworth however the proposal does not outline the relevant details of this multi-faceted work. This control-related approach which is about fusing different information types in general and the related feedback looks, does not provide the necessary details to build motivation systems for Hyperon/PRIMUS as it only speaks about these aspects on a high-level from a human cognition perspective. AI systems demand more detailed formalizations and implementations which would have demanded a corresponding description in the proposal which goes beyond what Frankel and Bedworth proposed.

  • Expert Review 2

    Overall

    5.0

    • Compliance with RFP requirements 5.0
    • Solution details and team expertise 5.0
    • Value for money 5.0
    A very solid proposal which would fulfill the requirements of the RFP in an elegant and systematic way

    The proposed dynamic-control/motivational framework seems interesting and applicable to Hyperon as well as NN-based AGI systems, and explicating and prototyping it as proposed here seems valuable...

  • Expert Review 3

    Overall

    5.0

    • Compliance with RFP requirements 4.0
    • Solution details and team expertise 5.0
    • Value for money 5.0

    Solid and interesting proposal largely based on the Human Information Processing work of Frankel and Bedworth. It is different from our current OpenPsi approach as it appears to be more focused on information processing rather than motivational drives. That being said, the proposers make a point of substituting the word motivation for emotion with the HIP literature. Coming at the question from a somewhat different perspective compared to our current work could prove to provide some new insights.

feedback_icon