The
Q-AL Assistant™ Architecture White PaperThe
Q-AL Assistant™ Architecture White Paper
Abstract
This white paper describes and explains the Q-AL Assistant Architecture and its relationship to the patented Digital Life-Form Simulation™ (DLF™) technology architecture, for which the Q-AL Assistant Architecture architecture is the interface. The description and explanation is one level deeper than that in our FAQ. The paper continues with a description of how to integrate the DLF and Q-AL Assistant interface architectures into a system that simulates life like processes, including simulated human intelligence as a form of Virtual Consciousness™ with natural language understanding, and this paper explains some key points about system deployment. The last section explains how to license the Q-AL Assistant technology architecture and the technical support program that is planned for it.
Note: For complete "how to build one" details, see our book How to Simulate Consciousness Using a Computer System.
Technology Architecture Descriptions
There are two technology architectures that make up the Q-AL Assistant technology architecture system: The DLF technology architecture is for simulating the underlying processes to generate life like behaviors and the Q-AL Assistant architecture is the Quasi™ human interface persona for communicating with humans using Virtual Consciousness.
Digital Life-Form Simulation Technology
Given what is currently known, there are only three ways to create a life like computer system: 1) Attempt to extract human knowledge and common sense and put it into a huge database (this has been attempted by Doug Lenat with the Cyc system and others, 2) include humans as a type of processor in a hybrid AI computer as with our own Q-AI Expert™ architecture design, or 3) Design a system that mimics biology and human conscious processes closely enough that it can function as a form of Virtual Consciousness, which can simulate learning, knowledge acquisition, and "common sense" on its own by mimicking say 5-10% of human conscious capabilities, as we have done with the Q-AL Assistant architecture design.
The Q-AL Assistant architecture rests on the premise that computer technology cannot have the attributes of a life-form, such as conscious behaviors, based on a purely mechanistic design. The reasons for this are beyond the scope of this document (see our book, referenced above, for details), but suffice it to say that state of the art Artificial Intelligence (AI) and Artificial Life (AL) design strategies are oversimplifications of the reality of life processes because they ignore teleology (the study of goal-directed action), and the fact that telologic is the essence of the processes of all living systems. State of the art systems also ignore the fact that consciousness is a natural, limited, proactive, identification process (neither mystical nor transparent), and that the human concept formation process is quasi-mathematical. Thus the state of the art design strategies preclude systems that are good at mimicking life-forms from the outset.
While our Q-AL Assistant architecture is no more complex than a modern operating system, it provides a software layer for state of the art computers that adds simulated goal-directed behavior (in order to simulate life processes on a mechanistic system), plus two additional layers for Virtual Consciousness functions (sense perception and concept formation with natural language functionality). These three teleological software layers operating together make proactive, self-sustaining, self-regulating processes, such as the simulation of conscious intelligence, possible using ordinary mechanistic computer systems as an amimation platform. The essential idea is to mimic the way that the teleological processes of life-forms "run" on ordinary mechanistic chemical and physical processes. To accomplish this goal, the "teleologic interface layer" changes the identity of the computer system, and hence its causal capacity. In other words, it changes the way the computer system as a whole interacts with the world around it to enable it to better mimic biological behaviors in a way state of the art systems cannot. And while such simulations can never equal biology (precisely because they are not alive), they can mimic some biological processes, behaviors, and causation, provided they are designed with the appropriate identity. The layered model below shows this system design in the form of a table. Note: The functions in each layer of the table are analogous, not equivalent.
Biological Life-Forms
Digital Life-Forms
Layer 7
Conceptual Consciousness (Reason)
Simulated Conceptual Consciousness
Layer 6
Perceptual Consciousness
Simulated Perceptual Consciousness
Layer 5
Goal-directed Cellular Processes
Simulated Goal-Directed Behavior
Layer 4
Mechanistic Cellular Processes
Digital Life-Form Simulation Program
Layer 3
RNA, Protein, ATP Synthesis
Object-Oriented Prog. Environment
Layer 2
DNA Processes
Computer Operating System
Layer 1
Electro-chemical, Physical Processes
Computer Hardware
There are three fundamental reasons why state of the art computer and robotic systems are do not have Star Trek* like computer capabilities today:
1) All state of the art computers do is run passive scripts (programs) that work like falling dominos.
2) They do not sense reality directly (or do so only as pixels, not as objects like people). Most systems use only data made up by programmers and stored in text files.
3) They cannot calculate their own concepts based on sense perceptions. The only "concepts" they have are formed by human programmers and stored in text files.
The DLF technology architecture overcomes all three of the deficiencies indicated above, and one purpose of this white paper is to explain how it does so in more detail than was appropriate in our FAQ, but less detail than our book How to Simulate Consciousness Using a Computer System.
Point 1) Goal-Directed Action
Unlike biological systems, computer systems are simply mechanistic automatons. They are passive automations of human goals and actions that cascade through preprogrammed scripts like falling stacks of dominos. Computer automations are human goals and actions in digitally recorded form that no longer have the ability or motivation to act proactively and independently the way life-forms do, precisely because they are no longer part of a living organism. (That is, they are no longer part of a self-powered, self-regulating, self-programming, goal-directed system.) A DLF acts much more like a biological life-form than state of the art AL systems do. Because a DLF's "life" and "death" is simulated as part of the system telologic (layer 5 in the table above), DLFs must act independently to "survive" (like biological life-forms, rather than state of the art system designs). That is, DLFs are designed to be self-powered, self-regulating, self-programming, and goal-directed. Like real life-forms, DLFs can act, but in order to exert the simulated "effort" to do so, they must first "survive" by finding and "eating" simulated food to generate Energy Packets (EPs), which are then used to "power" their future actions. (Dead life-forms cannot act. Survival is a complex causal process: Life causes both its own continuation, and as long as life actions continue, it can cause other actions that may not necessarily be survival related. The following image depicts this complex form of causality graphically. The goal of survival must be met on each cycle for the process to continue, and as you can see, it is very different from the way mechanisms work.
(See reference 3 in Appendix A Section 1.2.1 of our book.)
If the life process stops, it not only can no longer cause its own future action, it can no longer cause any action. (And in real life-forms, it cannot be restarted.) As we pointed out earlier: Dead life-forms cannot act. In biology, survival involves continuously maintaining certain conditions (whatever a life-form's life requires), and that is why life is said to be "conditional" (Ayn Rand®): If the conditions are not met (such as air, water, and so on for mammals), life processes stop, the life-form dies, all actions cease, and the life-form disintegrates. The primary goal of life, therefore, is survival, and all other goals are in support of it because if survival is not achieved, life-forms quite literally go out of existence. This is why goal-directed behavior is central to maintaining the conditions required by life processes, why it is more complex than mechanistic causality (though it is based on it). And this is why teleology must be taken into account in order to create a realistic simulation of Life-Forms with Virtual Consciousness using a computer system.
Most actions of Life-Forms are related to their survival in some way. In fact, the so-called "higher" functions such as perceptual consciousness in mammals and reason in humans evolved precisely in support of survival. These behaviors do not exist separately from Life-Forms (in spite of unscientific claims by mystics), but rather behaviors such as consciousness and reason are attributes of Life-Forms, attributes that are primarily biological. To create a good simulation of intelligence that mimics biology (except for very narrow domains like playing chess), one needs to simulate consciousness. But in order to simulate consciousness, one needs to simulate the teleologic of goal-directed behavior. And, in order to simulate the teleologic of goal-directed behavior using a computer system, one needs to create an interface to adapt the complexity of goal-directed causality to the mechanistic nature of the computer, because mechanistic action is simpler in a causal sense, and the computer system's identity must be modified accordingly. (Note: A detailed discussion of this topic can be found in Chapters 2 & 3 of our book How to Simulate Consciousness Using a Computer System.)
Point 2) Simulated Sense Perception
State of the art computer systems do not sense reality directly as collections of objects like people and higher animals do. (They "see" only pixels or bitmaps.) To illustrate why this is important, and in light of the discussion of survival in the previous paragraphs, let us consider the following example: Imagine yourself in India walking through the jungle. You hear a noise behind you and turn to look to see what it is. How would you prefer to see whatever is there? Would you like to see what you would see as a human being? That is, as an object with attributes that are easy to recognize, such as the following image?
Or, would you prefer to see the object making the noise you hear like a state of the art computer would "see" it, as a collection of numbers like the following image?
Imagine trying to decide if the object is a threat or not from that list of numbers, instead of the seeing the tiger as an object with a few clear and obvious attributes!
Yet pixelmaps and/or bitmaps are all state of the art computer systems "see." To take a simpler example, for a computer a closed figure graphic of a "triangle" is also just a list of pixels. You or I see the triangle shape as attributes (like triangle "#1" shown below in the first image) and as an "object", whereas a computer "sees" it as the list of numbers (shown in the second image below):
Note: The shapes in the following graphic were hand drawn on a computer screen with the mouse to generate these data, and the list of pixels is for triangle 1. You and I see the triangle as an object at the conscious level, and that is what I am referring to, not the neuro-physiological level.
For complex objects there can be thousands, even millions of pixels. The Virtual Consciousness of a Digital Life-Form (DLF) can "sense" these pixels too (they are its data, the initial content of it simulated consciousness), but then it does further processing to identify the implicit attributes that distinguish the triangle as a foreground object from its background. As a result, the DLF is then able "see" the triangle as an object with a unique list of attributes (properties and associated measurement values), attributes the DLF calculates for itself. Performing these calculations is how a DLF simulates sense perception, a method that produces results that are analogous to the attributes a person would identify when perceiving a triangle. For large, complex objects, the process amounts to a form of data compression because thousands or millions of pixels can be reduced to a short list of attributes, which can then be used for further processing or converted to a graphic image for communication with people. But the main benefit is that the DLF performs its own real-time, firsthand, identifications of the world, unlike state of the art AI systems which depend on the second hand identifications made by the consciousness of the programmers who predefined them and store them in file for the computer to process. (Note: A DLF does not "see" a graphic like the image of the tiger or triangle, but a short, invarient attribute list that captures its identity and that is used for processing purposes. The is analogous to and simulates the invarient objects people see. The attributes of objects "seen" by a DLF have been extracted from the sensory input. No one knows exactly how the human brain does the same thing, how it identifies and "stores" the images that people see and use for further processing, but it is very unlikely that they are stored as tiny pictures. As what we know of how the brain operates increases, the list architecture described here will be replaced by our Reality Identification Cortex (RICX) that will extract attributes more like the visual cortex does so. But the method is irrelevant. What is relevant is that the attributes of an object's identity are extracted in some form, so they can be used by processes farther up the processing hierarchy, such as those of simulated emotions or concept formation.)
Shown below is an actual example computed by our prototype demonstration program of a simulated percept of triangle #1 that a DLF calculated from the pixelmap shown above. The DLF Program interface was used to draw triangle #1 on the screen with the mouse. The simulated "sensing" algorithms then extracted the pixels just as a state of the art computer system would to simulate "sensing," but the processing did not stop there as it does with state of the art systems.
The DLF Program also processed the pixel information to identify the properties and measurement values (the attributes) of the objects in the scene. The pixel information for triangle #1 was transformed into a list of the property and value pairs for each of the three lines that constitute the triangle as an object. Taken together, the lists for the three lines are the identity of this specific triangle.
We can say the DLF that performed these processes has simulated the perception of triangle #1 because, as a human would, it identified the triangle as an object, as an integration of attributes. This is what we mean by simulated sense perception, and the next graphic shows the actual results.
Future versions of the DLF Program will produce a composite list, instead of a separate list for each line in the triangle, and the result will be a smaller "footprint" for the identity of each object. Obviously, there is not much data compression in this example because triangle #1 does not have very many pixels. However, for large colored objects the economies of scale will be significant for a DLF that can identify large, complex objects with a short list of attributes. (Which properties are extracted will obviously affect the size of the lists, but these choices can be tailored to specific applications.)
The main advantage of the data compression is not for storage, but rather for economies in processing and communicating information about the objects that the DLFs have identified. To compare objects a DLF has "perceived," for example, instead of having to process many lists of potentially millions of pixels, only relatively short lists of properties and values need to be processed. And, since the original pixels are stored as well, additional properties can be extracted later depending on the needs of a given application for the technology.
Note: The property list below is similar to one for common Web software objects in that it is a list of property/value pairs. The classification methods explained below could as easily be applied to XML Web objects, for example, as the objects shown here.
If you look at the objects in graphical form in the top image and their identities in the lower one, you can see the differences between the identities of a triangle and a circle. These are simple examples, but more complex shapes can be identified in a similar way. The bottom line is that this approach enables a DLF to "identify" not only the objects in its world, but also the relationships that are implicit between those objects. Once the object identities are "known" to the system as lists of properties and measurement values, there are an enormous number of relationships that can be identified, and one set of them is the similarities and differences that can enable DLFs to calculate their own abstract concepts.
Point 3) Calculating Concepts
State of the art computer systems cannot calculate concepts as abstract identifications of perceived objects and communicate meaning using ordinary words and sentences in human languages like people can. (Computers only store and manipulate text and graphic pixels as symbols for human information and knowledge, not their own.)
Once a large number of simulated percepts have been computed for various objects in a DLF's world, a DLF can calculate simulated concepts (abstractions) using these data, symbolize the concepts with words provided by a human tutor (such as the word "triangle"), and then communicate using ordinary human languages about these and other objects the DLF "sees." But exactly how are such concepts calculated?
The process is actually fairly simple (Source: Ayn Rand, see ref. 2, our book, Appendix A Section 1.2.1 ):
If you look at the objects in the image above, as a person you will see similarities and differences. Next, if you compare them and make attribute lists for objects 1, 3, and 4 (like the lists for triangle #1 above), you will see that the attributes you identify for the triangles share all the same properties, and differ only in their measurement values, (which are the entries on the right side of the lists above). This relationship comes from the fact that the triangles share the same shape measurement range, even though the specific measurements are different in the identity of each unique triangle. On the other hand, the circle and the square, which serve as the foil in the comparison, do not share all the same shape measurements (the square's identity is not shown), so we can say that the triangles are similar as opposed to the other shapes, which are different. Based on this identification, we can group the triangles together due to their shared measurement ranges, treating each as a unit or a member of the group of two or more similar members, as opposed to other objects with different identities that do not share the same measurment ranges.
Obviously, this process becomes more effective with a much larger sample of objects, but it works even with only two objects, and concepts can always be updated as the context of examples expands. (In fact, updating concepts as context changes is a requirement to accuracy.) The measurement range that results from this process can be left as a calculation or can be translated into an English definition with the help of a human tutor such as: "A closed shape consisting of three straight lines connected at their end points" when sufficient concepts have been formed to do so.
Finally, the simulated concept formation process is completed when DLF symbolizes the group of units with the word "triangle" (provided by a human tutor), and then it has successfully calculated a concept. Within the context of the objects we compared, we have an objective concept because it is calculated from direct, firsthand "observation" of objects outside the DLF in reality, concept's definition is its calculated measurement range for triangles, and any object that has measurements within that range can be identified as a triangle, even triangle #5, though we did not use it in our comparison. Triangle #5 can be recognized even if it had not been "seen" before by the DLF because it fits the definition for the concept, it shares the same measurement range as other triangles, its identity is similar to them.
Simulated concepts formed using this process are a new kind of data structure that does not exist in state of the art computer systems. The simulated concept is a data structure that can enable a simulated Life-Form to store large amounts of information about the world in a manner analogous to the way humans do using conceptual abstractions. This capability can make human to computer communication much easier and productive.
To gain significant advantages, the Virtual Consciousness of a DLF must repeat the concept calculation process over and over for every different type of object it encounters, for all the important relationships between objects, and connect these concepts to the natural language words its human tutor provides the words for it. The tutor would also help the DLF calculate higher level abstract concepts like "closed shape," "animal," "plant," or "object." The result is a hierarchy of concepts as shown below. (Note: In computational epistemology, a similar hierarchy of Web objects classes would be called an "ontology.")
A DLF's concepts are objective, as opposed to the subjective concepts used in state of the art AI systems. These latter are subjective and arbitrary because they are defined by human programmers, not an AI system itself. This is an important difference. Once thousands of simulated concepts have been formed, they reduce processing by many orders of magnitude as a DLF "identifies" objects and relationships in the world around it with its simulated consciousness, but this is only a useful strategy if the DLF has objective concepts that actually correspond to that world, connected to reality by a DLF's calculation chains.
For state of the art AI systems, every circle or any other object they "see" is different from the last one, just another pixelmap or bitmap like the ones shown earlier. Even the simulated percept of an object, such as the one shown above for the triangle, is unique to that object, and such identity lists can refer to only one object at a time. On the other hand, with the concepts it calculates, a DLF needs only one small symbol, the word "triangle," to identify any triangle it encounters past, present, or future. The word, through the concept calculation chains that gives it meaning by connecting it to objects in reality (including all their relationships), identifies the overall contextual relationship for the DLF for triangles. This capability represents an additional form of data compression because it makes it possible for a DLF to refer to all that it "knows" about triangles with a single symbol, a symbol that is a common referent for both DLFs and people (the word "triangle"). Collections of conceptual symbols, with their associated meaning, can then be processed together to represent abstract ideas.
At this point in the explanation of DLF simulation technology, we have explained enough details for some of the benefits of simulated sense perception and concept calculation to become apparent. The way DLF simulated consciousness leads to artificial intelligence is that it enables DLFs to "identify" and "recognize" objects in reality, for themselves, rather than second hand as arbitrarily defined by human programmers.
In the graphics shown below, the graphic on the left shows how a DLF can "identify" specific objects by their identity as simulated percepts (attribute lists), lists that are self-calculated, internally by DLFs. This process is called simulated perceptual identification, and it is how a DLF simulates perceptual consciousness. The process is analogous to what we observe in animals and people. Objects are recognized as objects, not pixel lists or bitmaps, so a DLF "sees" the world sort of like we do. It "sees" the world as simulated objects, rather than pixelmaps.
The graphic on the right below shows how a DLF can use conceptual identification, in other words, the calculation chains of its concepts, to "identify" many more abstract relationships in reality than it could without using chains of simulated concepts. In fact, it is just this capability that enables a DLF to communicate with people using ordinary words and sentences.
To understand how a DLF can decode and encode simple human language sentences, consider the image again on the right above, then note the levels of abstraction in the DLF's memory that result from the calculation chains of its concepts. The graphic above shows the percepts, concepts, and relationships to reality that a DLF would able to decode after it was trained by a human tutor. The following example presupposes the DLF has "perceived" a large number of objects, "observed" them in various relationships, and calculated concepts that enable the DLF to identify abstract relationships such as "shape," "object," "world," "existence," and so on. The following example explains how an English sentence can be decoded and another one encoded by the Quasi interface working in conjunction with the DLF Simulation Technology, and based on the image of "Reality" in the graphic above. A human tutor might ask the Quasi DLF:
Quasi, "Is the circle nearer to the oval or the rectangle?"
The process the Quasi DLF performs is: Perceive the sentence and the objects in reality, do a conceptual recognition, calculate the answer for the relative positions of the objects, encode another sentence containing the answer, and output the new sentence to the tutor.
Quasi says: "The circle is nearer to the oval."
Here is a more detailed explanation of how the Quasi DLF can do this:
First, using the process that simulates his consciousness (his ability to identify), Quasi "perceives" the question as a collection of text objects. Next, assuming he has neutral simulated feelings at that time (no conflicting goals), he makes the simulated choice to parse the sentence by tracing the calculation chains for each word/concept pair the sentence contains, and then decodes its meaning by tracing similar conceptual calculation chains for the grammatical and syntactic information contained in the word order and punctuation of the sentence.
Once Quasi has identified the meaning of the question after several simulated conscious events in his Virtual Consciousness, he "perceives" the shapes in the reality, identifies their locations relative to each other using the location measurements in their attribute lists (their identities). Quasi then compares the actual positions of the objects to the meaning he has calculated for the question in earlier conscious events. This result, which is already in his memory from his previous actions, shows Quasi that the circle is nearer to the oval than the square. Based on this new identification, Quasi then reverses the process he uses for decoding sentences to calculate a new sentence, and then outputs the sentence to communicate the answer he has identified to the human tutor who asked. Each of these steps would require many simulated conscious events to complete, many passes through what we call a C.Event loop. (See Chapter 3 and Chapter 5 of How to Simulate Consciousness Using a Computer System for complete details on this process.)
The process of communicating using simple sentences is a very complex activity (but no more complex than a modern computer operating system in terms of simulation system programming requirements). It is also a good example of what it means to simulate consciousness as a process of the identification of reality, as a form of Virtual Consciousness and for a computer system so designed to communicate with humans using natural language. A DLF's capability to "observe" reality and communicate using ordinary sentences is what offers the most value to our future customers because it will finally enable people to work with computers and robots in many of the ways they do with other people---and in a way that is similar to how the characters do it on the Star Trek TV series.
While the DLF Simulation Technology Architecture simulates life processes and conscious functions for DLFs, all these processes still need an interface persona for making communication with humans easier. The Q-AL Assistant architecture is really just the Quasi interface applied to DLF simulation technology.
As is explained in the white paper for the Q-AI Expert™ architecture, the interface personas can be graphics and text like the mock-up shown above, or more complex interactive video personas for when DLF technology is running on computers, PDAs, augmented reality systems, robots, or the Internet. The degree of sophistication and complexity for the Quasi interface can be determined and configured as needed.
Obviously, before any design work begins on the DLF, its Virtual Consciousness, and Q-AL Assistant integrated system for any specific customer application, a detailed needs analysis should be performed to determine a snapshot of the current system in the domain of the project for the organization planning to implement the Q-AL Assistant Architecture system, an identification made of the desired state of the system when the project is complete, a determination made of the changes required to migrate the system from the current state to the desired state, and a detailed design plan written that specifies both how the migration will be accomplished and how the final system design will be deployed and integrated into the wider domain of the organization as a whole. For the Q-AL Assistant architecture in particular, a risk assessment for using the technology must be made. DLFs simulate "choice" to select some of their actions, and DLFs are therefore capable of error. This fact, that self-programming systems can make errors as they learn, must be taken into account by customers.
Before DLF Simulation Technology can be deployed, the Q-AL Assistant Developer's Kit product needs to be created and tested, as well as product documentation and training developed. This is necessary so each of our future customers do not have to teach basic concepts to the DLF system, but rather we can have our own engineers do so one time only, make multiple copies of the "educated" DLFs computer files, and then be able to offer our customers a starter system that already "knows" about many objects, abstract concepts, and relationships.
Note: The Developer's Kit is currently in the planning stage and on hold pending funding.
Once the objective of producing the Developer's Kit has been accomplished and customers identified, the procedures for deploying the Q-AL Assistant architecture in a given organization should be specified in a system design plan as is standard procedure for IT projects. Many of the procedures will obviously be specific to the organization doing the deployment. However, there will be a need for some general procedures in the deployment process. This last requirement is a consequence of the risks inherent in the fact that DLF Simulation Technology is goal-directed, self-programing, simulates "choice," and is therefore capable of "mistakes." (Unlike state of the art technology, where the only mistakes are made by human designers and programmers.) As a result of this fact, customers will need to be educated about the risks of "errors" on the part of DLFs during and after system deployment, as DLFs continue to learn about their world. We plan to develop training and support programs to help our customers deploy and operate their DLF simulation systems.
If you decide to license Q-AL Assistant Architecture technology, the first step is to email us so we can negotiate a license agreement. Once the license agreement is approved, we will work with your IT team to support them as they develop the design for your system. In addition, we can design a custom training program in parallel with your development project, so that training courses are available at the time your system is deployed.
At Blue Oak Mountain Technologies, Inc., we are looking for "angel" investors and corporate partners to help us take DLF Simulation Technology to the next level. If you are interested, just mailto:greg@blueoakmountaintech.com and tell us how you want to help.
< Back to DLF Simulation Technology page
© Copyright 1998 - 2008, Blue Oak Mountain Technologies, Inc. All Rights Reserved
* STAR TREK and related marks are trademarks of Paramount Pictures Corporation.
** Ayn Rand is a registered trademark of the Ayn Rand Institute
Welcome Page | About Us | Testimonials | Services | FAQs
