HOLISM VERSUS REDUCTIONISM FOR HUMANOID ROBOTS

ABSTRACT

All existing Brain/Robot/Human theories such as symbol manipulation or evolutionary computing are no more than powerful metaphors. New metaphors may be more appropriate to develop intelligent humanoid robots. This paper argues that it is more scientifically interesting and fruitful to evolve a society of robots rather than to program a robot. The high-technology industry, the internet, the quantum computer, the earth's ecology, and the theater are all powerful metaphors that can be used to build robots and their societies. Both holistic and reductionistic approaches may contribute to the creation of humanoid robots, and the proper balance between them is advocated.

I. INTRODUCTION.

1. Ideas and systems should be considered together with their whole environment, including history. Every scientific idea, even revolutionary, reflects the knowledge and beliefs of its time. There has always been an establishement of ideas (council of elders, Church, state, community of noted scientists) and revolutionary ideas that violated these paradigms. Both were useful and necessary from the point of view of the mechanism of emergence and acceptance of new ideas/systems. It was so, it is so, and most likely it will remain so in the years to come. Establishment, revolution, acceptance of revolution within the system, constitute a cyclic process, that has to be understood in order for us to be able to model it. This concerns also the brain/man/robot debate.
   
   
2. I believe the following:
   • A. In the past, theories of man, brain and robot, have been developed in parallel, influencing one another. It will remain like this, but the role of interaction and society in creation of intelligence will grow.
   • B. You cannot understand the human brain without the body, and the body without the brain. You cannot understand the man without the society, and the society without man. You cannot understand a complex system without understanding its emergence.
   • C. You cannot build Intelligent Robotics, Artificial Intelligence or Artificial Life without having to deal with the material problems of sensors, motors, wires, effectors and base technologies.
   • D. Computer is material and therefore subject to material constraints of speed, time, power consumption, area and complexity. When we build robots and create theories of their operation, we cannot neglect this fact.
   • E. In a long run, implementing AI/AL in physical robots will prove to be a better approach to understand intelligence and life, than constructing single-point intelligent behaviors; such as chess programs, software ALIFE simulations or automatic theorem provers.
   
   
   
3. In this text I will use the names "brain theory", and "robot theories" interchangeably. I will discuss some weaknesses of previous theories and, in this context, I will propose a new approach to robotics. I will understand this approach, however, as only one more metaphor for intelligent robot design, and not a statement about the nature of intelligence or life.

II. THE WORLD IS CONSTRAINED BY ITS OWN NATURE

III. REDUCTIONIST AND HOLISTIC THEORIES

• There exists an objective truth, which is the world (Universe) itself. Theory of truth has been precisely formulated by Tarski: ``when a sentence in a language agrees with the state of the world, it is true, when it does not, it is not true''. Therefore, one cannot say that "all theories are equally true/untrue". They are not. On the other hand, for more general theories this measure of agreement can be binary, multivalued or fuzzy. Besides its truthfulness, a theory is characterized by its completeness. It results from Goedel's work that these theories (such as arithmetics) cannot be proven true and complete at the same time. In this context, all robot theories are true to a certain degree. Thus, mathematically, any robot theory is incomplete. From the system point of view, a non-trivial theory cannot be complete because only the world (Universe) itself is isomorphic to the world. The theory can be true but useless, and can be untrue but heuristically useful, because every theory is only a brick in the building of science, built by many generations (like Newton's gravity theory is formally untrue but very useful). Theories of robot are thus no different from previous theories in physics and biology, but can be also evaluated the way the formal systems are.
  
• In the process of science, society and the evolution of technology, theories obtain feedback from the real world, and those that are not good will not survive as models of world. This way the Ptolomeus astronomy or Creationist Theory were doomed to fail in the long run, although they were successful for hundreds or thousands of years.
• Holistic theories were more true, but reductionistic theories had higher heuristic value.
  
  For example, La Mettrie's theory of a human as a machine with pulleys, gears and steam was not true and extremely reductionist. He was not saying that this is only an analogy. He meant it was literally true. But this theory advanced science more that the holistic theories of his time.
  
  
• In the same way, the evolutionary approaches to robotics - modern darwinisms - believe in their models literally and not as metaphors or analogies. In my understanding every theory formulated in a natural (not formal) language is true only by analogy, and not literally, because only the Universe is a true model of itself. Because both man and Universe are inifitely complex, every particular theory is a simplification. Thus, although robot theories are useful, they are limited. A dogmatic tendency to stick to one theory is hindering the progress.
  
  In history, theories are created in phases (like the succeeding mechanics of Aristotle's, Newton's, Einstein's, quantum, quark, etc.). There are links between the phases of development of theories:
  • Useful holistic theories of phase N+1 cannot be built without knowledge of reductionist theories of phase N
  • Holistic theories of phase N+1 include holistic theories of phase N
  • The reductionist theories of phase N+1 should include the reductionist theories of phase N, because we understand now that a more complex system is able to model simpler system (like a Turing-class computer can model a mechanical system, or a quantum computer can model a standard Turing-like computer.) Whether the next phase theory includes as subset or negates completely or selectively the previous theory, it is not created in abstraction from this theory.
• When I write "is able to model" I state that we have to take into account computational complexity, Goedel-like and Heisenberg-like constrained principles. Thus we cannot say that "a computer algorithm can, in principle, exactly solve the graph coloring problem" for arbitrary graph with 100,000 nodes because it would physically require a computer with more memory cells than atoms in the universe, and more time than the Universe exist. This is the difference of material and ideal, discussed by St. Thomas. Platonic theories that separate ideal world of numbers and abstractions from an imperfect world of matter, have no chance to be true in robotics, although they advance our thinking. The Turing machine with indefinite tape is thus a platonic concept, and, not of-this-world, idea. It is of limited use in understanding how this world works. And thus not helpful to build a real physical robot. This was a typical mistake of Turing and successive reductionists, from which they had subsequently to withdraw and correct their statements.
  These kinds of physical constraints, embedded in systems, were not taken into account by the reductionist theories. Ironically, the atheistic hard-AI-believers such as Minsky or Simon, were more platonic than the aristotelian St. Thomas and traditional thomists. Ironically, the materialists create theories that do not take material laws into account.
  Thus the theory of robot should be reductionist, holist and material. It should be based not only on ideal thought processes but on physical interactions, processes and phenomena. Artificial Intelligence theory based on building material robots in the long term will be closer to the truth than the theory that assumes use of ideal beings such as recursive functions operating on infinite memories.
  

IV. THE WEAKNESSES OF THE REDUCTIONIST THEORIES

• A) Human is a mechanical/hydraulical machine,
• B) Human is an electrical machine,
• C) Human is a cybernetic machine (Wiener, Ashby),
• D) Human is biological machine,
• E) Human is physiological/chemical machine,
• F) Brain is a computer (Turing Machine).

• E) Brain is a Quantum Computer,
• F) Brain is a network of computers (Finite State Machines),
• H) Brain is a Evolvable Hardware such as an Field Programmable Gate Array (FPGA),
• I) Brain is an Internet - Internet is a Brain,
• J) Holographic theory of a brain.

V. BRAIN THEORIES AS ANALOGIES AND METAPHORS

1. Thus, we can safely say that none of the brain theories is true, and all are true to a certain degree. The reductionists do the same mistake as religions have done in their early phases; to treat what is symbolic, literally; and what is analogous, as a one-to-one mapping.
   Regretfully, the reductionists are not able to recognize their mistake. I did not see this idea of ANALOGY in writings of Minsky, Simon, De Garis [DeGaris93,DeGaris97,DeGaris00,Buller98], Moravec and other hard-AI-believers. When I read their books I have the impression that they truly believe that IT IS SO AS THEY WRITE.
   Observe, however, that this dogmatism is indeed their strength. As it happens, the people who try to see all aspects and understand from all sides are very slow to reach conclusions. Meanwhile, one who looks briefly and speaks quickly, a.k.a. the reductionist, can make an impact with more speed.
2. The advocates of the reductionist theories have strong appeal to public with their catchy simple ideas. They have therefore a strong influence in a short run. Such theories are easy to explain and thus have some appeal, especially to an unsophisticated mind (nazism, communism, primitive churches and sects, advocates of primitive interpretations of cybernetics or darwinism). In a long term they cannot win, because you cannot explain the complex system by a reduction to a simple system. If they were right, the human and the nature would be finite, and thus the real progress in science would be soon stopped, because all questions would be answered (the "end of science theory"). And with this they cannot agree. Again, locally in time and space these theories may have positive impact, and even wars and sufferings caused by them are non-zero sum games and may be necessary elements of humanity growth [Wright99].
3. When applied to Robot, all current "hot" theories such as Genetic Algorithms, Evolutionary Computing, intelligent agents, Neural Nets, Symbol Manipulation, Fuzzy Logic modeling, brain modeling, "brain building", despite the much influence they have now, will share the fate of former reductionist theories, but will still remain useful components in the evolutions of science, technology and human society.
   
   

VI. THE WEAKNESS OF HOLISTIC THEORIES

1. Now that we criticized the reductionist theories, let us observe that holists have their own sins.
2. If a theory is too general and too holistic, it can be understood by very few people and it tries to accomodate too much to make any point. Telling everything truthfully, it tells nothing of use or of interest. By trying to make no mistakes, it avoids telling some local truth that may be useful. By trying to avoid bias, no learning can be accomplished. Every learning process involves certain bias and hence the learning without bias is not possible. Induction is nearly always false ("all birds fly") but is the main way to learn. Holistic theories tend to concern themselves with observing phenomena and stating facts and rarely trying to explain the phenomena. If they do try to explain, it is usually not very constructive. It must be asserted that holistic theories rely on a very passive paradox. They are often collections of obvious and unexciting truisms. The God of holism truly needs the Satan of reductionism to make the world. The infinite cannot be explained without the finite, nor the complex without the simple.
   
3. With these irrationalities and pragmatic impediments the great literature and the holy texts of many religions took another approach to tell the all-encompassing truth - that of paradoxes and contradictions. The solution was to tell stories that can be understood more or less metaphorically. The story of Original Sin is here a perfect example because it allows for so many interpretations, and each of them quite creative.
   
   
4. It is extremely difficult to create a complete holistic and constructive theory of everything (and a theory of robot is theory of everything!).
5. Such theory would be necessary to build truly humanoid robots - the "spiritual robots" (we distinguish here between humanoid robots that certainly will be build and will exceed humans in many areas, and the philosophical concept of "spiritual robots" [Kurzweil99] as a new form of life).
6. Because it seems very unlikely to create such a theory, we are left with two discourses. One, allow the robot science to remain in the realm of the dialectically understood reductionist/holist loop of theories. Two, free the robot science from the loop in favor of a theatrical interpretation: one that offers insights through metaphor, drama and allows the observer any number of creative interpretations. We do not know what will be evolved, but we will intentionally create an environment in which the mystery of creation in narrower, theatrical sense, can happen.

VII. LIFE AND INTELLIGENCE. NEW METAPHORS TO BUILD ROBOTS.

1. The fundament of life is reproduction and survival, including competition (for space, for food, for female). Thus, no true humanoid robot can be created that would be not able to reproduce freely in a real world environment. If we believe in evolution, let the evolution create such robot, otherwise we are creating robot for us and not for Universe, so we are not creating true intelligence. Because science did not (yet?) create a technology that would allow for real reproduction, and we model robots without their real need for survival, we are not working on true models of life yet. Cyborgs will be humans with protheses, it will be not really a new life form. Only if we would create life from scratch on nano-technology level (Drexler) it would be a true emergence of life. Everything else is a simulation. One may figuratively say that we are cheating in our competition with God, because He created humans using "his own ash" and we try to use His ash). Only if Earthly "genotype seeds" would be send to another planet, and would create life, would we be able to talk about creating life and intelligence in a philosophical sense. But this is still science fiction.
   
   Let us then take another approach.
   
   
2. Much of human race's current efforts to model brain and build robots are just plays similar to a theatre. Theater can be great and deep, it tells much about life and world, but it is not the world itself. Theater is a good metaphor. Primitive religions and societal powers originated from theater, so theater is a natural way to express symbolism outside purely material means of comunication. It is the oldest art and the source of symbolic thinking. Reconstructing the emergence of human society cannot be done without understanding the theater. The role of theater was recognized by many great thinkers, anthropologists, and theater theorists/reformers [Campbell, Elliade,Stanislawski, Grotowski, Kantor]. Greek science and philosophy were preceded by hundreds if not thousands years of mystery plays and theatre. We will especially concentrate on great myths of ancient cultures, such as the myth of Prometheus. Theater is at the origin of all civilizations and is easier to model by robots than the sexual reproduction or the "survival of fittest" between human races or societal organizations. Interestingly, one of the first books ever written on theater, by Hero of Alexandria, as early as in the first century, was devoted to a robot theater [Hero-of-Alexandria]. "Interactive radio" was also predicted by great theater reformer and director Bertold Brecht in a book "Radio Theory", in late twenties [Brecht67]. Brecht wrote about a transformation of broadcasting from distribution only to a communication system in which the listeners actively influence the contents of the action. But he was not able to predict the Internet technology of today. Ramon Lullus; the medieval priest Anzelm, philosopher-teacher of Saint Thomas; the rabbi from Prague; Pascal; Descartes and Leibnitz; they were all fascinated by robots, mechanical puppets, Golems, mechanical men and talking heads. There is a long-term link between robotics and theater, creation and mystery, and this relation has never been just for entertainment, or only accidental. Time has finally come, that it can be investigated in its fullest.
   
   
3. Here I will propose a new theory for a robot, understanding that it is only a one more hypothesis in a long chain of theories that will be as long as humanity - a growing and emerging system by itself - will exist.
   
   The presented theory is based on four analogies:
   • J) Robot as a High-Technology industry,
   • K) Robot as an Internet of Quantum Computers,
   • L) Robot as an Earth ecological system, the world (Gaia-like hypothesis),
   • M) Society of Robots as a Theater.
   
   
4. In contrast to other researchers, I do not treat these analogies literally, just metaphorically. My claim is only heuristic, and I believe that only a practical success verifies the theory, and only locally.
   
   Because every theory is useful only locally in time and in its application area (the most successful computer/robot applications were based on very limited principles - Deep Blue chess program, Samuel's checkers program, ping-pong robot, etc.), a better theory is the one that allows to create better limited robots in a given moment of time. Not one that creates unverified general claims. Ultimately, every real robot will include a system of many theories.
   
   So, the Genetic Algorithm theory is in no way "philosophically better" than for instance the heuristic search theory. They are both models, and one of them can be locally better to model some particular behavior of a robot.
   
   I am not a purist, I am a pragmatist and I do not believe in any particular theory for building robots. My goal is to take metaphors from the world to build interactive plays/games for a robot theatre/society. In the past our research group took methods from Logic Synthesis and applied them to Data Mining, being part of a robot [Perkowski99a, Perkowski99c]. I believe that the science and world are full of analogies, all of them could be useful if just the robot researchers would find time and interest to study them.
   
   
5. Let us now explain first the analog methodologies (models, theories) listed above, and next how they will be used in the Oregon Cyber Theatre, our reductionist/holist robot model.
   
   
   • Referring to point J. Although may be Nature uses Darwinian algorithm, human society has invented another methods of solving problems, so the darwinists cannot exclude that other learning processes may be emergent in Nature. Mathematics, physics, logic, Search theory or game theory give better problem-solving algorithms than the Genetic Algorithm in many practical problems such as deriving formulas from examples. Why then should we be restricted to Darwinian evolutionary approaches only? For instance, the modern high technology companies and high-technology world market are the most complex systems that ever existed. Let us observe how a new microprocessor chip in Intel, the most complex system ever build by humans, is constructed. In my opinion, nobody with common sense would propose to develop such a chip using Genetic Algorithm or search methods. Engineers and researchers in "design sciences" developed many specialized theories of optimizing layout, logic, chip architecture, routing, circuits, etc. Each of them requires highly sophisticated knowledge of mathematics or/and physics. It would be totally hopeless to build such chip based on any single theory of mind, that the dogmatic purists believe are the base of everything.
     Modeling the way Intel designs chips would help us build a robot brain. Developing theories, creating prototype software, testing, verifying, prototyping, doing this everything with very many local and global feedback loops. There are many models of the outside world. How do we know that the Nature does not work like this?
     Therefore we proposed [Perkowski99a,Perkowski99c] to use logic-synthesis/evolvable-hardware/FPGA-design-methods as a competitor to GA and NNs to design systems that will learn in real time. I am not excluding GAs (Darwinian, Baldwinian, Lamarckian, etc.), I just want to find a local, proper place for evolutionary methods in the whole framework of ideas for humanoid robots.
     
     
   • Referring to point K. In the "The Society of Mind" theory, Minsky proposed perhaps for the first time a powerful metaphor of a brain as a society of individual agents [Minsky]. These ideas were next proven practical by Rodney Brooks [Brooks], and become now dominant in robotics. I accept this metaphor in its entirety, and in addition I propose to use the analogy to the Internet with its distributed control and self-growth mechanisms. Because even the entire Internet cannot solve NP-complete problems of useful size, I assume that in future the individual computer nodes of Giga-Net will be quantum computers. With the very inexpensive microcontrollers, sensors, memory chips and Field Programmable Gate Arrays (FPGAs), this theory can become practical soon. Building a robot with 100 microcontrollers, each controlling a single muscle, already becomes a reality even for a university with average funding. Because we cannot build quantum computers yet, we will model their constrained and probabilistic behavior in FPGAs and microcontrollers, of course sacrificing much speed and computational performance, but learning their nature and possible applications.
     
     
   • Referring to point L. The above remarks relate also to the "robot as a world" metaphor. Combining the above two metaphors with other system-theoretical models and data mining systems, we will be able to create models of learning and behavior more powerful than the existing one-sided models (NN, FL, GA, GP, search, game theory, symbol manipulation, automatic theorem proving).
     The problem, unsolved so far by anybody, is only this - how to combine different models? Much recent research is devoted to this subject, but so far no systems have been created that would demonstrate solving this dilemma. I believe that the combination methods should use adaptation, learning, voting and negotiating processes, game theory and self-emergence, and be thus "evolutionary", but not necessarily based on current evolutionary paradigms.