Skip to content

Latest commit

 

History

History
495 lines (431 loc) · 25.8 KB

README.md

File metadata and controls

495 lines (431 loc) · 25.8 KB

Sensory Atomese

This repo explores how perception and action within an external world might work with the AtomSpace.

TL;DR: Explores philosophical approaches to perception & action via actual, working code using low-level AtomSpace sensory I/O Atoms. A software conception of "basal cognition". The experimental lab for this is "perceiving" filesystem files, "moving" through directories, and likewise for IRC chat streams.

Philosophical Overview

The issue for any agent is being able to perceive the environment that it is in, and then being able to interact with this environment.

The goal in this project is to find a theory of a minimum viable low-level sensory API. Having such a theory would clarify how learning systems might learn how to use it, or, more generally, how to "use things", how to have a cause-and-effect in the universe.

For OpenCog, and, specifically, for OpenCog Atomese, all interaction, knowledge, data and reasoning is represented with and performed by Atoms (stored in a hypergraph database) and Values (transient data flowing through a network defined by Atoms).

It is not hard to generate Atoms, flow some Values around, and perform some action (control a robot, say some words). The hard part is to (conceptually, philosophically) understand how an agent can represent the external world (with Atoms and Values), and how it should go about doing things. The agent needs to perceive the environment in some way that results in an AtomSpace representation that is easy for agent subsystems to work with. This perception must include the agent's awareness of the different kinds of actions it might be able to perform upon interacting with the external world. That is, before one can attempt to model the external world, one must be able to model one's own ability to perform action. This is where the boundary between the agent and the external world lies: the boundary of the controllable.

Traditional human conceptions of senses include sight and hearing; traditional ideas consist of moving (robot) limbs. See, for example, the git repo opencog/vision for OpenCV-based vision Atomese. (Note: It is at version 0.0.2)

The task being tackled here is at once much simpler and much harder: exploring the unix filesystem, and interacting via chat. This might sound easy, trivially easy, even, if you're a software developer. The hard part is this: how does an agent know what a "file" is? What a "directory" is? Actions it can perform are to walk the directory tree; but why? Is it somehow "fun" for the agent to walk directories and look at files? What should it do next? Read the same file again, or maybe try some other file? Will the agent notice that maybe some file has changed? If it notices, what should it do? What does it mean, personally, to the agent, that some file changed? Should it care? Should it be curious?

The agent can create files. Does it notice that it has created them? Does it recognize those files as works of it's own making? Should it read them, and admire the contents? Or perform some processing on them? Or is this like eating excrement? What part of the "external world" (the filesystem) is perceived to be truly external, and what part is "part of the agent itself"? What does it mean to exist and operate in a world like this? What's the fundamental nature of action and perception?

When an agent "looks at" a file, or "looks at" the list of users on a chat channel, is this an action, or a perception? Both, of course: the agent must make a conscious decision to look (take an action) and then, upon taking that action, sense the results (get the text in the file or the chat text). After this, it must "perceive" the results: figure out what they "mean".

These are the questions that seem to matter, for agent design. The code in this git repo is some extremely low-level, crude Atomese interfaces that try to expose these issues up into the AtomSpace.

Currently, four interfaces are being explored: a basic text-terminal, a single-file reader/writer, a unix filesystem navigator, and an IRC chat interface. Hopefully, this is broad enough to expose some of the design issues. Basically, chat is not like a filesystem: there is a large variety of IRC commands, there are public channels, there are private conversations. They are bi-directional. The kind of sensory information coming from chat is just different than the sensory information coming from files (even though, as a clever software engineer, one could map chat I/O to a filesystem-style interface.) The point here is not to be "clever", but to design action-perception correctly. Trying to support very different kinds of sensorimotor systems keeps us honest.

Typed Pipes and Data Processing Networks

In unix, there is the conception of a "pipe", having two endpoints. A pair of unix processes can communicate "data" across a pipe, merely by opening each endpoint, and reading/writing data to it. Internet sockets are a special case of pipes, where the connected processes are running on different computers somewhere on the internet.

Unix pipes are not typed: there is no a priori way of knowing what kind of data might come flowing down them. Could be anything. For a pair of processes to communicate, they must agree on the message set passing through the pipe. The current solution to this is the IETF RFC's, which are a rich collection of human-readable documents describing datastream formats at the meta level. In a few rare cases, one can get a machine- -readable description of the data. An example of this is the DTD, the Data Type Definition, which is used by web browsers to figure out what kind of HTML is being delivered (although the DTD is meant to be general enough for "any" use.) Other examples include Interface Description Languages, the X.500 and LDAP schemas, as well as SNMP.

However, there is no generic way of asking a pipe "hey mister pipe, what are you? What kind of data passes over you?" or "how do I communicate with whatever is at the other end of this pipe?" Usually, these questions are resolved by some sort of hand-shaking and negotiation when two parties connect.

The experiment being done here, in this git repo, in this code-base, is to assign a type to a pipe. This replaces the earliest stages of protocol negotiation: if a system wishes only connect to pipes of type FOO, then it can know what is available a priori, by examining the connection types attached to that pipe. If they are BAR+ or FOO+ or BLITZ+, then we're good: the or is a disjunctive-or, a menu choice of what is being served on that pipe. Upon opening that pipe, some additional data descriptors might be served up, again in the form of a menu choice. If the communicating processes wish to exchange text data, when eventually find TEXT- and TEXT+, which are two connectors stating "I'll send you text data" and "That's great, because I can receive text data".

So far, so good. This is just plain-old ordinary computer science, so far. The twist is that these data descriptors are being written as Link Grammar (LG) connector types. Link Grammar is a language parser: given a collection of "words", to which a collection of connectors are attached, the parser can connect up the connectors to create "links". The linkages are such that the endpoints always agree as to the type of the connector.

The twist of using Link Grammar to create linkages changes the focus from pair-wise, peer-to-peer connections, to a more global network connection perspective. A linkage is possible, only if all of the connectors are connected, only if they are connected in a way that preserves the connector types (so that the two endpoints can actually talk to one-another.)

This kind of capability is not needed for the Internet, or for peer-to-peer networks, which is why you don't see this "in real life". That's because humans and sysadmins and software developers are smart enough to figure out how to connect what to what, and corporate executives can say "make it so". However, machine agents and "bots" are not this smart.

So the aim of this project is to create a sensory-motor system, which self-describes using Link Grammar-style disjuncts. Each "external world" (the unix filesystem, IRC chat, a webcam or microphone, etc.) exposes a collection of connectors that describe the data coming from that sensor (text, images ...) and a collection of connectors that describe the actions that can be taken (move, open, ...) These connector-sets are "affordances" to the external world: they describe how an agent can work with the sensori-motor interface to "do things" in the external world.

Auto-wiring and theorem proving

The goal of describing system blocks with some DTD or IDL language, and then hooking them up "automatically", is not new. Four approaches are worth mentioning.

  • ProLog and Answer Set Programming (ASP). In these systems, one asserts a collection of facts/assertions, and then says "let's go". Prolog uses a chaining algorithm, and ASP uses a SAT solver to determine a "solution" satisfying the constraints embodied in the facts. One does not get back the actual inference chain that was used. For this project, we want to know how things got hooked together. We're not interested in a true/false satisfiability answer, but rather in finding out how assemble a processing chain.

  • Automated theorem provers. These are systems that, given a set of facts/assertions, together with a collection of inference rules, provide an actual inference chain, aka a "proof", that explicitly gives the steps from the hypothesis (the input claims) to the result (that the input claims are consistent and simultaneously satisfiable.) This gets closer to what we want: we want the chain. However, unlike theorem provers, we're not particularly interested in satisfiability, other than to find out if a partially-assembled pipeline still has some missing end-points. This is why we work with sheaves or jigsaws: we want to know not only how to assemble the jigsaw pieces, but to also know what the remaining, open (unconnected) connectors are. Theorem provers do not provide such info.

  • The SOAR Cognitive Architecture. This is a production system that applies production rules to state. This captures a different aspect of what we want to do here: we want to have data, and to apply rules to that data to transform it. However, our data is not so much "state" as it is a "stream": think of an audio or video stream. SOAR selects a rule (analogous to an inference rule, in theorem proving) and applies it immediately, to mutate the state. By contrast, we want to think of inference rules as jigsaw pieces: how can they be assembled? Once assempled, then these rules can be applied to not just "state" in a single-shot fashion, but repeatedly, to a flowing stream of data (say, video, to find all cats in the video).

  • Programming language compilers. These are able to take high-level specifications and convert them into an equivalent program written in assembly language. The assembly instructions can be thought of as jigsaws, and when they are assembled, the mating rules must be closely followed: the output registers in one instruction must attach to the inputs of another. The resulting program has to be runnable, executable. This is exactly what we want here: an assembly of interconnected jigsaws. However, unlike a compiler, our "assembly language" consists of various abstract processing components: filters, transforms, etc. It doesn't run on a (real or virtual) CPU, but on an abstract machine. It is not registers and RAM that the instructions/jigsaws act on, but on sensory data (again: think video/audio). Worse yet, the instructions aren't even fixed: new ones might get invented at any time. These might supplement or replace old ones. Compilers also want a program, written in a high-level language, as input. In this project, we won't have such a program; our situation is closer too SOAR or ProLog or theorem provers: we have a collection of jigsaws (insturction) to assemble, but no high-level program to specify that assembly.

The above systems solve some of the aspects of what we want to do here, but only some of them, and not in the format that we actually need. The above should give a flavor of why we're embarked on the crazy journey we're on. No one else does this.

Autonomous Agents

The sensori-motor system is just an interface. In between must lie a bunch of data-processing nodes that take "inputs" and convert them to "outputs". There are several ways to do this. One is to hand-code, hard-code these connections, to create a "stimulus-response" (SRAI) type system. For each input (stimulus) some processing happens, and then some output is generated (response). A second way is to create a dictionary of processing elements, each of which can take assorted inputs or outputs, defined by connector types. Link Grammar can then be used to obtain valid linkages between them. This approach resembles electronics design automation (EDA): there is a dictionary of parts (resistors, capacitors, coils, transistors ... op amps, filters, ...) each able to take different kinds of connections. With guidance from the (human) user, the EDA tool selects parts from the dictionary, and hooks them up in valid circuits. Here, Link Grammar takes the role of the EDA tool, generating only valid linkages. The (human) user still had to select the "LG words" or "EDA parts", but LG/EDA does the rest, generating a "netlist" (in the case of EDA) or a "linkage" (in the case of LG).

What if there is no human to guide parts selection and circuit design? You can't just give an EDA tool a BOM (Bill of Materials) and say "design some random circuit out of these parts". Well, actually, you can, if you use some sort of evolutionary programming system. Such systems (e.g. as-moses) are able to generate random trees, and then select the best/fittest ones for some given purpose. A collection of such trees is called a "random forest" or "decision tree forest", and, until a few years ago, random forests were competitive in the machine-learning industry, equaling the performance seen in deep-learning neural nets (DLNN).

Deep learning now outperforms random forests. Can we (how can we) attach a DLNN system to the sensori-motor system being described here? Should we, or is this a bad idea? Let's review the situation.

  • Yes, maybe hooking up DLNN to the sensory system here is a stupid idea. Maybe it's just technically ill-founded, and there are easier ways of doing this. But I don't know; that's why I'm doing these experiments.

  • Has anyone ever built a DLNN for electronic circuit design? That is, taken a training corpus of a million different circuit designs (netlists), and created a new system that will generate new electronic circuits for you? I dunno. Maybe.

  • Has anyone done this for software? Yes, GPT-4 (and I guess Microsoft CodePilot) is capable of writing short sequences of valid software to accomplish various tasks.

  • How should one think about "training"? I like to think of LLM's as high-resolution photo-realistic snapshots of human language. What you "see" when you interact with GPT-2 are very detailed models of things that humans have written, things in the training set. What you see in GPT-4 are not just the surface text-strings, but a layer or two deeper into the structure, resembling human reasoning. That is, GPT-2 captures base natural language syntax (as a start), plus entities and entity relationships and entity attributes (one layer down, past surface syntax.) GPT-4 does down one more layer, adequate capturing some types of human reasoning (e.g. analogical reasoning about entities). No doubt, GPT-5 will do an even better job of emulating the kinds of human reasoning seen in the training corpus. Is it "just emulating" or is it "actually doing"? This is where the industry experts debate, and I will ignore this debate.

  • DLNN training is a force-feeding of the training corpus down the gullet of the network. Given some wiring diagram for the DLNN, carefully crafted by human beings to have some specific number of attention heads of a certain width, located at some certain depth, maybe in several places, the training corpus is forced through the circuit, arriving at a weight matrix via gradient descent. Just like a human engineer designs an electronic circuit, so a human engineer designs the network to be trained (using TensorFlow, or whatever).

The proposal here is to "learn by walking around". A decade ago, the MIT Robotics Lab (and others) demoed randomly-constructed virtual robots that, starting from nothing, learned how to walk, run, climb, jump, navigate obstacles. The training here is "learning by doing", rather than "here's a training corpus of digitized humans/animals walking, running, climbing, jumping". There's no corpus of moves to emulate; there's no single-shot learning of dance-steps from youtube videos. The robots stumble around in an environment, until they figure out how things work, how to get stuff done.

The proposal here is to do "the same thing", but instead of doing it in some 3D landscape (Gazebo, Stage/Player, Minecraft...) to instead do it in a generic sensori-motor landscape.

Thus, the question becomes: "What is a generic sensori-motor landscape?" and "how does a learning system interface to such a thing?" This git repo is my best attempt to try to understand these two questions, and to find answers to them. Apologies if the current state is underwhelming.

Self-observing systems

Perception need not be limited to "the external world"; one may also observe oneself. Action need not be limited to the movement of limbs; it can also be a control over one's own thoughts. Imagine the case of "stewing in one's own juice's": a dreamlike state, where you ruminate over old memories of pst events. As one does so, one selects, picks and chooses: fond memories are noted for their emotional content, are given further thought; boring memories stay unexamined, unless forced.

The perception-action system described above can be aimed not just at the external world, but also at internal state. Imagine, for example, a trained LSTM or maybe an LLM that is placed into a "dream state", where it generates a sequence of free-association outputs. An agent can watch over this stream, and, in response to certain outputs, it can "prompt" the system, to guide further "reminiscences". More directly, it can "choose" to force the system to concentrate on a specific topic. To "focus one's thoughts". This kind of hierarchical layering, where an agent steers the thoughts of an underlying system results in a form of self-awareness and self-control.

Thus, in addition to the previously-described perception-action agents (e.g. traversing the file system or interacting via chat) one can build an agent that monitors the state of a neural network, and then controls it, via prompts, or perhaps much more directly with gates (tanh/signmoid blending.)

Related ideas

A distantly related set of ideas can be found in the SOAR Cognitive Architecture from Laird, Newell & Rosenbloom. Examples of SOAR agents can be seen in the SOAR Agent github repo. SOAR is a production rule system, and so resembles the AtomSpace query subsystem. The manner in which SOAR rules are applied resembles the OpenCog Unified Rule Engine (URE) (which is unsupported and now deprecated) and also OpenCog Probabilistic Logic Networks (PLN) (also unsupported & deprecated).

There are many differences between this work and Atomese:

  • SOAR production rules are written in ASCII (in the SOAR language), and are stored in flat files. AtomSpace production rules are written in Atomese, and stored in the AtomSpace.
  • SOAR production rules are applied to "state", and work within a "context". This is similar to the AtomSpace, which can store state. However, the AtomSpace can also store the rules themselves (as "state") and the rules can be applied to external data streams (e.g. audio, video).
  • SOAR production rules are crafted by humans, encoding knowledge. Atomese rules are meant to be algorithmically generated and assembled.
  • SOAR state mutation is boolean-valued: either something is done, or it isn't. There does not appear to be any concept of Bayesian possible-worlds.
  • SOAR appears to use a very simplistic forward-chaining approach to inference. For any given SOAR state, a collection of possible rules is determined. Of these, one rule is selected, and then it is applied to mutate the state. It has long been recognized that other kinds of chaining is interesting: not just forward chaining, but also backwards. One might not just chain (as in ProLog), but ask for constraint satisfaction (in ASP and automated theorem provers.)
  • The OpenCog URE and PLN elements are capable of performing chaining. They are currently deprecated/obsolescent, for a variety of technical and philosophical reasons.

A nice, quick & easy overview of SOAR can be found here: "An Introduction to the Soar Cognitive Architecture", Tony Kalus and Frank Ritter (2010)

Status

Version 0.3.0 -- Experimental. Basic demos actually work. Overall low-level parts of the architecture and implementation seem ok-ish. The upper-level parts have not yet been designed. The grand questions above remain mysterious, but are starting to clarify.

Provides:

  • Basic interactive terminal I/O stream.
  • Basic File I/O stream.
  • Prototype Filesystem navigation stream.
  • Prototype IRC chatbot stream.

The Architecture Overview provides a more detailed and specific description of how the system is supposed to look like, and how it is to work, when it gets farther along. The Design Diary documents the thought process used to obtain code that actually works and does what it needs to do.

See the examples directory for working examples.

The AtomSpace Bridge provides an API between the AtomSpace and SQL. It almost conforms to the system design here, but not quite. It should be ported over to the interfaces here.

The Vision subsystem provides an Atomese API for OpenCV. It is a proof-of-concept. It should be ported over to the interfaces here.

Design Overview

The Atomese agent framework needs to have some way of interacting with it's environment. Obviously, reading, writing, seeing, hearing. More narrowly: the ability to read a text file in the local file system. The ability to read directory contents, to move throuogh directories. The ability to behave as a chatbot, e.g. on IRC, but also as a javascript chatbot running in a web-page. Also possibly running free on twitter, discord, youtube.

The goal here is to prove the very lowest layers, just the glue, to convert that stuff into Atomese Atoms that higher-layer Atomese agents make use of to communicate with, interact with the external world.

The Architecture Overview provides a detailed description of how this can work. A general overview can be found in the AGI 2022 paper: Purely Symbolic Induction of Structure.

General system architecture is discussed in a number of places, including the various PDF and LyX files located at:

See also:

Design specifics

Details of the design in this git repo are explored in several places:

Build and Install

This git repo follows the same directory structure and coding conventions used in other OpenCog/AtomSpace projects. This cannot be compiled before installing the prerequisite AtomSpace. So build and install that first.

Then:

mkdir build; cd build; cmake ..
make -j
sudo make install

Examples

See the examples directory. The simplest example is for pinging text between two xterms. Other examples include opening, reading & writing a single text file, navigating the file system, and a basic IRC echobot.

It will probably be useful to read the Architecture Overview first.

Important All of this is pre-alpha! These examples are too low-level; the intent is to eventually automate the process for hooking up sensors to motors. Basic design work continues. But for now, these show some of the low-level infrastructure; the high-level stuff is still missing.