SVG logo by Gaëtan Tapon (Paris, France)

RefPerSys, a free software symbolic artificial general intelligence project

  RefPerSys (voir refpersys.org) est un projet (en 2020 à 2022, bénévole) de recherches et un logiciel libre en intelligence artificielle, inspiré par les travaux de feu Jacques Pitrat (1934 - oct.2019). Voir la journée d'hommage à J.Pitrat. Le lecteur est supposé capable de lire l'anglais.
Nous recherchons une source de financement possible (via des applications), par exemple non-exhaustif via l'Agence Nationale pour la Recherche - en s'intégrant dans un montage de projet -, ou un projet innovant collaboratif BPI France ou HorizonEurope or ECS collaboration tool ou CLAIRE. Veuillez alors contacter Basile Starynkevitch par courriel vers basile@starynkevitch.net ou basile.starynkevitch@cea.fr, et diffuser cette page A4

  RefPerSys (см. сайт refpersys.org) свободное программное обеспечение (ПСОК под лицензией GPLv3), для системы Linux/x86-64, в области симболицеского искусственного интеллекта (рефлексы, самоанализ, метазнание ...) по идеям Jacques Pitrat мы ищем партнеров в контексте www.indiarussiainnovate.org это работа в стадии разработки. А также ищем потенциальных приложений и пользователей. Пожалуйста, свяжитесь со мной, basile@starynkevitch.net (Василий Дмитриевич Старынкевич), который плохо читает и плохо говорит по-русски, но свободно владеет английским и французским языками. Jacques Pitrat был директором моего докторского жюри (в области искусственного интеллекта) защищенного в 1990 году (в Университете Paris 6).

  RefPerSys (see refpersys.org) is a free software research artificial intelligence project and open source software (work in progress) for Linux, inspired by the works of the late Jacques Pitrat (1934 - oct.2019). See his blog on bootstrappingartificialintelligence.fr and read his book Artificial Beings, the Conscience of a Conscious Machine (ISBN 978-1848211018). We are looking for potential applications, partnerships, collaboration and some funding for working on this project. European R&D projects and calls for proposals like ITEA4 or 0HorizonEurope or ECS collaboration tool or CLAIRE comes to mind. Contact Basile Starynkevitch by email to basile@starynkevitch.net (home) and basile.starynkevitch@cea.fr (office, at CEA, LIST).

RefPerSys - refpersys.org

An acronym for Reflexive Persistent System. This is an ambitious research project (still unfunded, in 2019 - 2022 it is a hobby project; also very risky!), done on spare time with an open science mindset and ethics.

Early work in progress

We have in 2019 - 2022 a lot of ideas, and some code for Linux; in april 2021: 42KLOC measured by Wheeler's sloccount software. That code is on gitlab and preferably on github.com/RefPerSys/RefPerSys and github.com/RefPerSys/refpersys-in-c. Done by passionate developers on their spare time. Volunteers wanted; please send an email (introducing yourself) to basile@starynkevitch.net (near Paris, France, homepage here) and to team@refpersys.org for more.

The RefPerSys development team

The following persons are working (in 2019 - 2022) on RefPerSys on their spare time from their home and are the copyright owners of the embryonic software.

design ideas

• a mostly symbolic artificial intelligence long-term project, with ambitious Artificial General Intelligence goals.
• for Linux distributions (such as Debian, Arch, Ubuntu, ....), on x86-64 desktops or laptops; with funding and cooperation could be ported to embedded Linux systems
• free software, GPLv3+ licensed; some embryonic code (in C or C++, it should later become generated) available on gitlab.com/bstarynk/refpersys and mirrored on github.com/RefPerSys/RefPerSys (for experimentation of ideas).
• several ideas borrowed from Bismon (a work funded by Chariot and Decoder European H2020 projects), but not a single line of code. See also this draft report.
• reflexive, persistent, multi-threaded, with a heavily metaprogramming and bootstrapping approach (all its code -C++ files, its Makefile, etc...) should become generated by the RefPerSys system, à la Quine), aiming to offer a declarative, knowledge based language.
• The RefPerSys recurring insight is: progressively more and more expressive and declarative formalization of knowledge and of meta-knowledge (applicable to itself), combined with run-time reflection, self-profiling and introspection and meta-programming (notably generation of C++ code) approaches. RefPerSys should provide and contain its own sort-of IDE, using Qt or GTK and will -later in 2022- progressively incorporate other existing open source C++ libraries (perhaps TensorFlow, Parma Polyhedra Library, Gudhi etc...), we hope to give more or less declarative high-level know-how formalizations related to these libraries (with some implementation inspiration by the past and obsolete GCC MELT project). More and more C++ or other source code of RefPerSys should become generated (and re-generated again from even more declarative formalizations).
• We hope (in 2022) to have declarative rules and metarules with variables. Your help is wanted. A rule should have un unordered set of conditions and trigger a sequence of actions.
• In spring 2021 (e.g. around git commit `4303b8f1f8ced36`), we are using GNU readline for the initial REPL interface
• In summer 2020, we are strongly considering to give up Qt and switch instead to the simpler and lighter FLTK 1.4 GUI toolkit. See ongoing work in our git fltk-branch, which we hope to merge soon to our master git branch.
• heavily inspired by the pionneering work of Jacques Pitrat (1934 - october 2019, France). Read notably his Artificial Beings: The Conscience of a Conscious Machine book (ISBN 978-1848211018) and see his blog.
  In a few words, we dream of RefPerSys becoming (in a decade) some CAIA like system, but genuinely usable by a few people on Linux machines. Basile Starynkevitch followed Pitrat's courses, defended a PhD thesis under J.Pitrat's presidency and J-M.Fouet advisorship, and met J.Pitrat in person a few times a year during more than two decades, and has a february 2016 snapshot of CAIA (which is unfortunately difficult to use, since without documentation; see also github.com/bstarynk/caia-pitrat). The Malice system is a predecessor of CAIA (by the same J.Pitrat).

A draft document giving some design ideas is downloadable as refpersys-design.pdf. We are improving it in 2019-2022.

The currently temporary existing (in 2019 or 2021), on gitlab code base is a bit disappointing (since in summer 2020 most of it is still hand-coded by us, not yet generated), even to the Refpersys team. But we have a lot of fun doing it. It is constantly improving, publicly available, still unreleased, α-stage. We are aware of the replication crisis. We hope to generate most of RefPerSys C++ code in 2023. The external libraries used by RefPerSys may evolve with time, with the hope to provide some partial semi-formal description of their properties to the system. We are aware of Rice's theorem.

The major insight of RefPerSys is to generate most of (and hopefully all) its C++ or C code from higher level representations, based upon a cyclic graph of objects in virtual memory. That entire graph, called the persistent heap, is loaded at startup, and dumped at shutdown, from textual files (and later, external databases). It is garbage-collected.

values and objects in RefPerSys

The persistent heap contains immutable values and mutable objects. Each value is practically speaking a 64 bits word, often a pointer (which can refer to an object in memory, or to a small memory zone holding some data, conceptually some tagged union). Values are hashed and ordered, and are a The nil pointer usually represents a missing value.
Immutable values include:

• tagged 63-bits integers
• boxed floating point numbers
• UTF-8 encoded strings
• sequence of object references
• finite and ordered set of object references. So membership test is in O(log n)
• immutable instances (containing, like objects do, a class, a set of attributes and/or a sequence of components, both being fixed)
• closures; the closed code is represented as an object (called the connective object), and the closed values (represented as a sequence) are arbitrary and could be nil.
• etc... (e.g. small vector of floats or of 32 bits integers, or JSON values)

• Their unique objid
• Their mutex lock
• The class of that object, itself an object reference
• The space of that object, used for persistence (e.g. all objects in the same space goes into the same file or database); it is an object reference
• The modification time of that object.
• A finite and updatable map of attributes, made of attribute entries. Each such entry has an attribute (itself an object reference) with an associated non-nil value. All attributes of a given objects are different.
• An updatable and resizable vector of components, each component being a value or nil
• optional pointers to C++ functions to apply closures or to get magical attributes.
• Some extra optional payload. Object payloads contain data which does not fit into the general ObjVLisp inspired model, e.g. GUI widgets, file handles, dictionnaries mapping string to values, etc...

Most (but not all) objects are persistent (so reloaded at startup and dumped to disk or databases at shutdown).

metaprogramming and introspection in RefPerSys

An important feature of RefPerSys is meta-programming and code generation. During its runtime, the multi-threaded refpersys process should generate C++ or C code, compile that code into some shared library, then dlopen that generated shared object, and get function pointers inside it using then dlsym. See also the Program Library HOWTO, and the C++ dlopen mini HOWTO and the manydl.c example. Generated C++ code could be temporary or persistent (in that later case, it is version-controlled).

Another important feature of RefPerSys is introspection coupled with runtime profiling (see also time(7)...) which is reasonably possible in generated C++ code with the help of the garbage collector (explicit GC-ed call frames being singly-linked lists) and/or Ian Taylor's libbacktrace library and the dladdr(3) function. During its runtime, RefPerSys can inspect most of its call stacks. We later hope to use machine learning techniques on the profiling data and get some dynamic optimization, tailored to the concrete problems approached and actual applications of RefPerSys.

declarative knowledge and bootstrapping in RefPerSys

The insight of RefPerSys is to represent high-level expert system rules, fact bases and machine learning artefacts as collections of objects. The C++ (or C) generator should be described by meta-rules compiled by itself to C++ (or C) code. So we are taking a bootstrapping approach, first generating C++ or C) code by hand-written C++ (or C) code, then replacing some of it by generated C++ or C code from higher-level knowledge representations, then progressively and incrementally improving these representations by higher-level and more abstract constructs. This should be worked on in 2022 or later.

In practice, RefPerSys requires some graphical user interface with autocompletion facilities, and displaying graphically small parts of its heap (perhaps as abstract syntax trees or nested boxes), inspired by the past Centaur system.. It is well known that GUI toolkits like Qt or FLTK are more easily used with C++ code generators, which are parts of RefPerSys. So we need to work a lot.

a rewrite in C effort was attempted (for political reasons) in fall 2021. See code on github.com/RefPerSys/refpersys-in-c.

talks

RefPerSys has been presented at a seminar organized by AFIA in Paris, on March 6^th, 2020, dedicated to the memory of Jacques Pitrat, who died in october 2019. Contact Basile Starynkevitch for details. Slides are here in PDF. The video of the talk in French is here.

Please suggest by email to team@refpersys.org any AI conference, meeting, or workshop where this project could be presented.

documents

• Draft of a paper (by Basile Starynkevitch and Abhishek Chakravarti) in French: Le projet RefPerSys, un succcesseur potentiel du système Caia de Jacques Pitrat - submitted in june 2021 to ROIA journal: PDF file (32 pages)
• The RefPerSys high-level goals and design ideas draft report (work in progress, 2019 - 2020, mentioned above).
• Jacques Pitrat's book Artificial Beings: the Conscience of a Conscious Machine, ISBN-13: 978-1848211018 (Wiley ISTE), is suggesting an interesting roadmap and provides an inspirational approach, justifying the interest of a reflexive and persistent system.
• A draft of a short hightlight paper (2 pages) submitted to ECAI2020 about the Development and Architecture of RefPerSys: a multi-threaded Reflective Persistent System for AI (november 2019; we are aware that this conference is very selective). That paper was not accepted, but one (out of three) of the reviews has been very encouraging, saying : The concept of the paper is good. However, the writing skill / style is not satisfactory.. That paper explains our roadmap and vision.
• Our wiki contain topics which are discussed and work in progress. Notice in particular the Personal data in RefPerSys wikipage, which might be indirectly related to European GDPR. However, be aware that we are not lawyers.

dependencies

The RefPerSys software depends on many other common stuff, notably a recent GCC compiler (in practice, GCC 10 in summer 2021), for C++17 with its standard library, a recent Linux kernel and C standard library (e.g. GNU glibc or Musl Libc), the JsonCPP library (for persistence) and the Qt or soon FLTK graphical toolkit, and the Onion HTTP server library, the libbacktrace library. The GPP preprocessor is useful for some documentation. The AddressSanitizer facilities of GCC are sometimes useful.

A multi-core x86-64 Linux computer with more than 16Gbytes of RAM and at least 100Gbytes of free disk space is recommended. Qt is being replaced by FLTK, more friendly to RefPerSys garbage collector (since Qt, unlike FLTK, is creating threads without notice).

download α-stage source code

The GPLv3+ licensed α-stage embryonic free software source code of RefPerSys is on gitlab.com/bstarynk/refpersys.
We practically are using git as a centralized version control system, because we are very few and sill don't understand well in 2019 or 2020 how a genuinely distributed version control system could play nice with a reflexive, orthogonally persistent, homoiconic, bootstrapped, self-generating Quine system.

Daily snapshot with source and binary

A daily snapshot (of the master git branch) could be downloadable on refpersys-snapshot.tar.bz2. It is usually but not always automatically generated every day, so could be broken or very buggy. That bzip2 tar archive contains both source files, some log files of build, and binary ELF executable refpersys for Linux/x86-64. It is made in good faith, but could be very buggy and might do wrong things if you run that executable (daily built on Debian/Sid/x86-64). So use the refpersys executable in that snapshot, provided with no warranty, at your own risk. For cyber-security reasons, avoid running ./refpersys as root.
Use ./refpersys --help to understand the program arguments and environment variables, if any (they would be conventionally prefixed RPS_ or REFPERSYS_, see environ(7) for more) that could be relevant.

It is preferable to download RefPerSys from gitlab.com/bstarynk/refpersys and look into the active git branches, then to build it on your Linux computer. The build time with make could last one or a few minutes.

Rewrite in C

For non-technical reasons, RefPerSys has tentatively being rewritten in C in october 2021 to spring 2022. That software (GPLv3+) is downloadable from github.com/RefPerSys/refpersys-in-c.

forum

A low-traffic forum hosted in France by FramaSoft (a non-profit association) is available on framalistes.org/sympa/info/refpersys-forum. Please subscribe if interested. Be polite and civil there. Only for public technical discussion in written English about implementation, architecture, or running issues of RefPerSys, including building problems (so not for philosophical discussions about AGI or singularity). If you have issues in subscribing, contact basile-freelance@starynkevitch.net by email. A typical message could have been this one. If you have built some RefPerSys executable and have tried to run it, that forum is the right place to ask help. Mention the output of the ./refpersys --version command in your message.

Thanks to Niklas Rozencrantz, a Slack channel for RefPerSys exists on refpersys.slack.com

needs

The RefPerSys project needs more volunteers to actively contribute to it, and welcome any person having tried to build and run it (even some intern student in programming, if he/she is willing to contribute). At last, some team members are severely under-equiped and would welcome any 4 to 16 cores x86-64, Linux capable, working, PC with 32 gigabytes of RAM and at least a 240 Gb SSD disk. Sponsorship for traveling costs could also be welcome, to facilitate face to face meetings. Opportunities to present this project in face to face meetings are desired, and more generally a limited funding is welcome (but we have no legal organization for that in 2021). Mention potential relevant funding opportunities (e.g. HorizonEurope calls) or any technical issues in building or running it to us by email. Contact team@refpersys.org for more.

Contributing to RefPerSys

If you want to contribute to this RefPerSys project, please compile it, and send the output of ./refpersys --version by email (in written English, UTF-8 encoded, HTML5 formatted) to team@refpersys.org and to basile@starynkevitch.net. Do introduce yourself (give your geographical location and country, education, experience in programming on Linux, past open source projects you contributed to, approximate age, and motivations) and explain the applications you have in mind, etc... Feel free to also give a mobile phone number.

potential applications

The RefPerSys project welcome potential future applications (if in return they provide some sponsorship, contributions, or funding). Since it can be viewed as a high-level declarative programming environment based upon a metaprogramming and at end of 2021 some machine learning techniques, we believe it may become useful... However, the priority in 2021 is to generate most (hopefully all) of the C++ code of the RefPerSys system. Contact team@refpersys.org.

Of course, any real-world potential concrete application of RefPerSys (e.g. health sector, smart cities, smart grids, smart web, perhaps robot planning, civilian autonomous vehicles or drones, etc....) would require some support and a little bit of funding. We all are excited by this project, and we all want to have some read-world application for it and expect such applications to fund partly our research work.

An incomplete list of future applications (assuming a budget of a few hundreds k€ or US$ and close cooperation with other software developers and domain specific experts during a few years, with computer access to some industry or domain-specific open data or proprietary data or communication in open format) might include:

• in the smart city vision, helping in optimizing energy, electric grids, communication grids, and public transport networks.
• in Industry 4.0 manufacturing plants, improving and monitoring the manufacturing processes
• in health care, assistance in modeling and simulating the human body. In particular, experts (biologists, physicians, virologists, public health autorities) fighting against the Covid 19 pandemic might benefit from RefPerSys (communicating with other existing software systems and hardware devices
and partly modeling that pandemic). Other bioinformaticssystems involve very complex software which might be monitored and communicate with some RefPerSysprocess.
• Organization of, and digital assistance to, specific human communities: within legal and ethical frameworks (in Europe, the GDPR), the set of members of a community or organization could be partly modeled as a set of RefPerSys objects representing some semantic network.
• Against climate change, RefPerSys might be used to coordinate and perhaps monitor very complex existing weather forcasting software systems (using finite element methods running on TOP 500 supercomputers.
• In software engineering of large software (as big as the Linux kernel or as the GCC compiler), RefPerSys might help junior software developers to understand the software architecture of these large code base, e.g. by partly modelling their control flow graph, or call graph, or navigate in a class diagram (see also the DECODER project) perhaps extracted by some GCC plugin, whose C++ source code might be (once some funding is available) generated by RefPerSys.
• In astrophysics, RefPerSys might coordinate and help instrumenting very complex stellar dynamics simulation software running for months on supercomputers. If the source code of these software is available, RefPerSys might generate some instrumenting code suited for them.
• large industrial facilities like nuclear power plants, or grids of wind turbines involve very complex simulation code such as CODE ASTER, CAST3M and are composed of thousands of networked hardware components, which could be simulated by RefPerSys objects (also interacting with existing software), etc.
• European organizations or like BDVA might become interested in RefPerSys or projects like EuHubs4Data could also be interested.
• Encyclopedical knowledge (e.g. Wikipedia...) might profit of RefPerSys since it is modelling (in a slightly more structured way) a semantic network, like Wikipedia does already. See also Cyc and BabelNet.

possibly relevant Horizon Europe calls

European Smart Networks and Services Joint Undertaking programme

see here for details.

Contact by email basile.starynkevitch@cea.fr and basile@starynkevitch.net or on mobile (Monday to Saturday, 8am to 9pm French time) +33 6 85012359 or by snail mail (at home) to
Basile Starynkevitch
8 rue de la Faïencerie
F 92340 Bourg-la-Reine
France

Be aware that any technical documentation will be produced (probably) using the LaTeX preparation system (and explicitly not any Microsoft Word processor) in PDF format.

We are aware that there is no silver bullet and RefPerSys is a software tool only. If interested by submitting a proposal on HORIZON-CL2-HERITAGE or smart networks and services joint undertaking calls, please contact by email basile.starynkevitch@cea.fr and basile@starynkevitch.net.

related projects

• The Tunes vision
• The Maru system
• The PhantomOS persistent OS

Please contact basile@starynkevitch.net to suggest related projects, and have an hyperlink to them. We would appreciate a reverse link.

research questions

Several research questions are unanswered, including:

• dynamically adapting symbolic knowledge representation to the current (holistic) state of the system. For example, a set of objects can be first represented as an unsorted array, then changed to some sorted array, and then change to some linked list, some red-black tree
• using machine learning techniques to change these representations
• dynamically computing the order of evaluation in expert system rules
• using meta-knowledge (expressed as rules) to generate code
• using memoization techniques
• driving and configuring garbage collection (GC) by rules
• generating GC code at runtime, optimized for the current heap
• generating serialization code
• using machine learning techniques to forget some data (opposite of memoization) and code
• etc...

Helping the RefPerSys project

You can help the RefPerSys project by emailing team@refpersys.org with your first and last or family name, email, geographical location, then:

• Compile the code and suggest applications and funding
• Contribute a few lines of code (under GPLv3+ license)
• Add some hyperlink (HTML code like <a href="http://refpersys.org/">RefPerSys.org</a>...) on your website.
• Mirror weekly this website and the source code, in particular the snapshot.
• Mention refpersys.org in any academic (peer-reviewed) paper.
• Email even officials or bosses inside your organization about refpersys.org
• Mention refpersys.org in call for proposals, e.g. to HorizonEurope
• Mention refpersys.org on high-traffic websites like StackOverflow or other forum. Some of us could be banned from StackOverflow, so ask them to unban us.
• Send us a postcard
• Send us books about computer science or Linux open source software
• If you write a book (on paper), mention refpersys.org in it.
• If you can print a leaflet about or mentioning refpersys.org at your work place, please do (and sign it)