For centuries mankind has been intrigued by human intelligence and often dreamed of reproducing it in a non-biological form. Scientists have studied it and tried to simulate it. Since the late 1940's with the invention of the electronic computer the dream has seemed more attainable than ever. This pursuit has been gathering even more momentum recently and today we find that research with this goal has split into several different yet overlapping approaches, known as: Artificial Intelligence, Cognitive Science, Robotics and Artificial Life. These lines of research have resulted in the development of software such as: neural networks, expert systems, agents, natural language understanding, speech recognition, and chess playing (Deep Blue) to mention just a few. However they have yet to achieve a level of general-purpose intelligence that the average individual would say is anywhere near equivalent to that of humans. Some projects have developed software that demonstrates human equivalent skills but only in extremely narrow subject areas or functional domains.
This book documents research that has the objective of reaching human levels of intelligence in software but takes quite a novel approach that belongs more in the fields of Cognitive Science and Robotics than Artificial Intelligence. The approach is to understand and model the underlying principles of operation involved in learning and thinking and then to develop software to perform these processes. This software is expected to learn as a child would and begin to think as an adult. The ultimate objective is to imbed this software in machines that will benefit mankind. This approach requires an analysis of the research that has primarily been done in the field of Psychology. You could say the approach is to reverse engineer those features of humans and animals that are of interest. The result is a model called Adaptron.
Adaptron is an information-processing model of an agent that learns and adapts to a given environment. It also includes features that provide it with the ability to think. It interacts with an environment via senses and response devices or actuators. The model includes mechanisms for attention, concentration, memory, recall, feelings and associations. It is a discrete model in that it deals only with symbols. These symbols represent the stimuli from senses, combinations of stimuli, emotional feelings, the output responses and ideas of things perceived. Adaptron is designed to explore an environment and seek out interesting and rewarding experiences. Conversely it is designed to avoid the repetition of boring experiences or punished actions.
But Adaptron is more than a working model; it is a complete system. It covers the spectrum from the abstract / conceptual to the concrete / physical. It includes a theory, architecture, a design and a functioning computer program.
Several philosophical ideas are fundamental to the design of Adaptron. These include intelligence, consciousness, modelling and survival. Chapter 2addresses these ideas.
This rest of the book is structured along the same lines that software applications are developed and documented for businesses.
- The business must be understood so business analysis is performed. The resulting business model describes the business processes and business things such as workers, tools, documents, products, customers etc.
- The requirements for the software application must be captured from the people that will use it. The result is a description of the subset of the business processes that the application must automate and what business things it must know about in order to work. The requirements are documented from the users’ perspective. They form a list of all the features and properties to which the software application must conform.
- The users’ requirements must be transformed into software requirements so system analysis is performed. The resulting analysis describes what the software system must know and do from an information processing perspective. The analysis must be clear, concise, complete and unambiguous and in a form that can be used for designing the software.
- Tests must be designed and written so that the final software can be tested to make sure it is satisfying all the users’ needs.
- A high level design to structure the software is needed so software engineering is done. The chosen software architecture describes all the parts of the software and how they interact so that the software will meet all the requirements.
- The detailed design of the software parts or modules must be performed. The module designs describe the software processes, data structures to be used, screens and reports as seen by the user.
- The software must be programmed according to the design. The result is the software source code. It is compiled into executable software for running in production.
- The tests are performed and the design and/or software are fixed until it works correctly and performs efficiently. This produces test results which show the correct operation of the software.
Chapter 3 of the book captures a wide spectrum of psychological facts, observations and evidence about animal and human intelligent behaviour to document an understanding of the subject. The study of psychology is the main contributor of this information.
Chapter 4 contains a statement of requirements that must be addressed. This is a subset of all human behaviour. It only includes requirements that I wished to achieve in this non-biological model. For example psychological theory says that one of the human drives is to reproduce and thus reproductive encounters are pleasurable experiences. This is excluded from Adaptron's requirements.
Chapter 5 is an analysis of animal and human learning and thinking. This analysis is my understanding and forms the theory of Adaptron. As a theory of learning and thinking it may well be wrong. However it is based on a significant volume of facts and observations in Chapter 3 and some reasonable deductions drawn from this information.
Chapter 6 describes what tests can be performed and the challenges involved in testing any artificially intelligent agent.
Chapter 7 contains the architecture and design of Adaptron. The architecture provides a framework or structure for mechanisms such as senses, actuators, attention and memory etc. to work together to achieve the whole. This architecture is comprised of a limited set of components that each performs a general-purpose function and that work together in a coordinated fashion.
Its design describes the process logic for algorithms that perform the functions of attention, memory, recall etc. that comprise learning. The design is based on a procedural programming language but can also be viewed from an object oriented point of view. Given a statement of requirements for a learning and thinking agent such as Adaptron there may be many different architectures and designs that work. The one described in this book is not necessarily the best possible design. At this stage of its evolution it is more important that Adaptron work correctly, not efficiently.
Chapter 8 contains the source code for Adaptron. It is a computer program written in Visual Basic.
Chapter 9 captures the results of several test runs of Adaptron in a variety of given environments.
Chapter 10 contains the conclusions.
Chapter 11 discusses further possible studies and the future of Adaptron.
As you read this book you will notice a distinctive lack of references. This is because it would take substantial work to add them and other authors and researchers have done an excellent job of compiling such lists. In Appendix A there is a list of recommended readings that all include extensive lists of relevant reference material.
I draw your attention to the glossary of terms in Appendix B that provides definitions for the terminology used in this book. The definitions are the understanding of these terms as they are used here. Most of the time this corresponds to the widely accepted meaning as found in psychology, cognitive science and artificial intelligence. However this is not always the case. As an example, the word 'punishment' is used to describe the flow of information that is negatively interpreted by an animal. It says nothing about how the information originates. It is not relevant for this theory whether punishment comes from another animal or from an event in the environment. According to this definition when you trip over a stone and hurt your knee the resulting pain is interpreted as punishment. Wherever a word is used with a narrower or broader scope than the usual understanding this is described in the glossary.
Correctly understanding the principles of operation of intelligent beings is one of the last substantial frontiers of science that has yet to be conquered. People have been observing and experiencing the workings of mental processes for thousands of years but have been unable to fathom the mysteries of their operation. It is obviously a complicated if not impossible task to perform completely. However scientists have been proposing theories on how the mind works for many years and many experiments have been carried out to verify these theories. [Ref's needed Stan Franklin’s Artificial Minds – include ACT-R V5.0] Often the problem with the psychological theories is they are described in insufficient detail to be reproduced in some other medium such as a computer program in order that they may be tested. Often the theories are far too narrow in their domain and thus do not give rise to generally useful systems or they are very specialized and break when applied to real life situations. Modelling learning and thinking has been the challenge and will continue to be the challenge for the next decade or two.
Adaptron is the result of my private research conducted over several years into simulating learning and thinking processes. This has consisted of substantial reading, discussions, observations and building of numerous computer models to simulate and stimulate ideas and theories about intelligent behaviour.
The objective of this book is to document the results of the research to date. This is done in a structured manner along the lines of the System Development Life Cycle which distinguishes between the 'what', requirements or goals and the 'how', design or means for their accomplishment. It proceeds from the more general and abstract, high level philosophy and analysis to the specific concrete low level of design and operation. The objective of this approach is to discuss the subjects of learning and thinking from different perspectives and at different levels of abstraction.
The primary objective of Adaptron is to understand the fundamental principles of operation of intelligent behaviour in order to duplicate it in the form of a computer program. Then the goal is to improve this program with more efficient, faster and robust algorithms. This is rather similar to how the aeroplane has evolved since its invention. Significant improvements were the invention or ailerons and the jet engine. A sub-goal of this research is to understand, model and duplicate the learning and thinking processes. Once such processes can be shown to operate correctly in a computer program they can be placed in all manner of agents which will result in the support, magnification and extension of mankind's abilities. For example, robots, traffic intersections and computers to name just a few can be made to act more intelligently.
The objective of the Adaptron design is to provide a functioning model of learning and thinking. Learning is modelled at the stimulus - response level that incorporates the components of attention and memory. This uses the processes of associative recall and concentration. Thinking is also modelled by incorporating attention and memory with the processing of recalled memories that become ideas or concepts. Emotional feelings are introduced as a fundamentally necessary component for the correct functioning of both learning and thinking.
Other objectives that have guided the design of Adaptron are described in more detail in Chapter 4.0 - Requirements. A summary of the primary objectives is to develop:
A model that is deterministic/predictable and symbol manipulating as opposed to one that uses weights or probabilities.
An agent that begins with no experience and no understanding of the environment into which it is placed but does have the ability to learn and think and knows what it likes and dislikes.
A general purpose agent that is not constrained to only operate in a small set of artificial environments.
A model composed of a limited number of integrated information processing components, each of which serves a different purpose. It should not be composed of a combination of programs or modules that each performs intelligently in some particular domain.
A primary driving force in this work has always been to consider an information processing perspective when analyzing and describing learning and thinking.
This model is intended to simulate/emulate learning and thinking in a deterministic environment with a constrained set of inputs and outputs. Learning and thinking are only two aspects of intelligent behaviour. Other aspects of intelligence such as pattern recognition and communication are only partially modelled and are out of scope of this book. Therefore Adaptron is a subset of what might be considered an intelligent agent. However that subset has been one of the hardest to model because learning and thinking are only indirectly observable. It can be studied only via the external behaviour (which includes communication behaviour) of an intelligent animal or through introspection of oneself, which is obviously subjective and thus prone to error. In humans we often detect the presence of intelligence via our ability to communicate with each other. This makes learning and thinking even more difficult to understand because the power of pattern recognition and the art of communication get in the way of making the required behavioural observations.
The two aspects of intelligence, Recognition (pattern, visual, verbal / speech, temporal) and Communication are out of scope of this model. They are only modelled to a sufficient level of detail to support the learning and thinking processes. Improvements to Adaptron in these two areas would add greatly to its capabilities and are areas for further research.
There are many other aspects of animal intelligence and mental activity that this book does not address. These are left for others to explain through extensions to or variations upon the ideas that are explained here. These Not-in-Scope subjects include:
- the soul
- sanity / insanity
- common sense
There are also several subjects that this model touches upon but in no way does justice to their explanation. These are also avenues for further study based on the Adaptron theory. The subjects with borderline scope include:
- pleasure and pain
- self awareness, free will, self control, consciousness
- pattern recognition
- subconscious action
This model does not contain any theories of how or where in the human nervous system and brain these functions are performed. Although there is mention of senses, the cortex and the cerebellum, their description and use are not to be considered scientifically accurate.
The theory of Adaptron is based on the assumption that if one can develop a sufficiently complete information processing theory of learning and thinking then one should be able to program a computer to simulate it. Such a simulation is then for all intents and purposes actually learning and thinking. This is based on the belief that learning and thinking processes do not require a biological brain or some metaphysical component for them to be performed. At this time there is no evidence to suggest these processes can or can not be performed by a mechanical, chemical or electrical machine.
The only kind of test that can be performed to correctly validate that Adaptron learns and thinks is an observational one. The behaviour of the simulation must be observed in a specific environment and compared to an agreed upon model of how an intelligent (able to learn and think) animal would behave in the same environment. It can not be determined by looking at the internal workings of the simulation. This is because, for comparison purposes, there are no correct or complete theories or models for how learning and thinking works. Also we can not inspect the internal workings of animals other than at the neurological level in order to model intelligent behaviour. This is currently impossible because the degree of complexity of the interconnections between neurones in the nervous system is extremely high.
However, should you feel that the theory of operation for Adaptron is similar to your own mental experiences you have some degree of proof for its correctness but only a non-scientific one. Testing for learning and thinking is explored in more detail in Chapter 5.0 - Testing.
This book makes no assumption about your background and experience. It has been written in practical terms for the majority of people to understand. However to understand the chapters on design and construction it helps to have a background in computer programming. An introductory knowledge of animal and human psychology and mental development would also be an asset. This book is not a textbook on psychology but does draw upon many psychological findings.