Search of a behavior emerging to to depart of the interaction of multiples entities through artificial life
Learning without previous knowledge
Manuel de la Herrán Gascón
- IA and Learning
- State of the art of the teorias of learning in how much to the knowledge minimal
- The theories Psychological
- State of the art in the models computacionales of learning
- How to learn
- Life artificial and systems emerging
Departing of the hypothesis of Newell and Simon on the systems of symbols physical, and of the posture reduccionista underlying in the systems that employ a architecture of levels independent as base for to build cognition artificial, it is analyze how to obtain the description of the characteristic that it must to possess a system cognitivo artificial for to permit the learning without knowledge previous, or to the less with the minimal knowledge possible.
The end last to the that it is guide this work is the definition of a system dynamical capable of to acquire knowledge in any environment of problem. In this line it is suggest the hypothesis according to the which the interaction of multiple entities with autonomy own that imitate the characteristic own of the beings live, within of a system cognitivo only, puede to produce a behavior emerging that originate the processes cognitivos superior, such as the learning.
[ Back to Index ]
IA and Learning
In the Artificial Intelligence (I) it is they can to observe, to large features, two approaches different:
- The conception of I as the attempt of to develop a technology capable of to supply to the ordering capacities of reasoning or discernment similar, or apparently similar to the of the intelligence humanizes.
- The conception of I as investigation relative to the mechanisms of intelligence humanizes, that employ the ordering as tool of simulation for the validation of theories.
The first approach is for it general the more practical, it is center in the results obtained, in the usefulness, and not so much in the method. In this approach it is straddle, for example, the Systems Expert. They are topics key in this address the representation and the management of the knowledge. The authors more representative of this approach they are McCarthy and Minsky, of the MIT.
The second approach it is guided to the creation of a system artificial that it will be capable of to accomplish the processes cognitivos human. From this point of sight not is so important the usefulness of the system created (what makes), as it is method employee (how it makes). As aspects fundamental of this approach it is they can to indicate the learning and the adaptability. Ambo present great difficulty for be included in a system cognitivo artificial. This direction is own of Newell and Simon, of the Carnegie Mellon University.
Is obligator to indicate that frequently both postures not it is they can to distinguish, neither even in many projects of the authors mentioned as meaningful in each a of they.
In the two positions, but especially in the second approach, one of the greater desires is the power to count with a architecture that support all type of process intelligent.
In the science cognitiva, the concept "Architecture" it is refer to the structure underlying to the dominance flexible of the process cognitivo, is to say, to the structure that sustain the processes cognitivos superior.
The architectures proposals as bases of the cognition humanizes it is designate Architectures Cognitivas, while that the corresponding for the cognition artificial they are calls Architectures for Systems Intelligent Integrated, or Architectures for Agents Intelligent, or Architectures General of Intelligence [Ros93].
The attempts of construction of systems cognitivos artificial it is bases in the hypothesis of Newell and Simon according to the which "a system physical of symbols constitute the middle necessary and sufficient for a action intelligent general" [New72].
If it is they could to explain the processes cognitivos superior of a way intrinsic, is to say, if it is could to demonstrate that the processes mental intelligent that accomplish the Man it is produce to a level superior with independence of the caps underlying that exist until the constitution physical of the entity intelligent, it is would demonstrate that is possible to create -through a system of symbols physical-, a structure artificial that imitate perfectly the mind humanizes through a architecture of levels, already that it is would to build said level superior through the combination of elements that not necessarily they have of be the neurons that form the level inferior in the human.
In change, if only it is they could to explain the processes cognitivos superior to level of the behavior neuronal, only it is would to imitate the intelligence humanizes through the construction of neurons artificial. For be exact, this affirmation it is conditioned for the certainty of the supposition (quite common) segun the which the neuronal is the more under of the levels relevant for the training of the the processes cognitivos. Arbitrarily, it is would have elected other level yet but under (molecules, atoms). Carried to the extreme, it is would reescribir the affirmation, sistituyendo "neurons" for "the more small particle of our universe", if this it would be discreet (not infinitely divisible).
The denominations "level superior" and "level inferior" they are arbitrational in how much to that it is puede to find with facility a level that it be yet more under that the that we have call "level inferior" -level of atoms- and it symmetrical respect to the level superior -conscience collective-. The existence of a conscience collective capable of it be communicated to a level superior to the of the individual seem evident in the studies on the behavior of some insects, always that we make the effort of not to interpret the term "conscience collective" from our point of sight subjective as individual. ¿How to obtain this? Not is dificil, if it is use a analogy decreasing a level: We imagine two cells (concretely, two neurons) of our body chatting amistosamente about of the possibility of that the joint of cells form a "conscience collective". The neurons they could to speak on this "conscience collective", but with difficulty they could to arrive to to understand it, not puede be a concept familiar for they. And.Or. Wilson, in "The insect societies" define the communication massive as the transmission of information, within of groups, that a individual particular not would to transmit to other.
The fact of to suppose that the behavior intelligent in the man it is produces to a level superior with independence of the levels inferior it is intimately related with the discussion between holismo or belief in that "the all is more that the sum of its parts" and the reducciónismo, or belief in that "a all puede be comprendído thoroughly if it is understand its parts, and the nature of its sum." [Hof87].
The efforts developed in Architectures General of Intelligence they are purely reduccionistas. For the opposite, the holismo underlying in the models conexionistas as the Nets Neuronales Artificial, suggest the aspect of the interdependence between levels, or it that is it same, the impossibility of to substitute a level (the connections neuronales, as system sub-symbolic) for other that accomplishes its same functions (system symbolic).
[ Back to Index ]
Drescher analyze with great wisdom the problem of the learning without knowledge previous [Dre91]. The first issue is to decide if exist or not knowledge previous, innate or to priority. The response to this question emerge in the discussion between the postures constructivista and nativista.
State of the art of the teorias of learning in how much to the knowledge minimal
- The posture constructivista or empirista, whose author more meaningful is Piaget, propose that the knowledge not is innate, but that it is obtain to to depart of the abstraction of the interaction with the world. Carried to its extreme, indicate that the experience is the only source of all the knowledge. The theories empiristas they were adopted for the psychology asociacionista.
- The posture nativista, characteristic of Noam Chomsky and Hering, it is center in to defend the existence of the knowledge innate or congenital.
Is important to indicate that this discussion it is produces exclusively in the dominance of the learning of the language natural. The great majority of the theories cognitivas they are basically constructivistas. Without embargo, many authors indicate the need of the existence of a knowledge innate, reconciling both postures.
Between they, Drescher assert: "if exist knowledge not innate, it must to exist a knowledge innate that it will be the knowledge on how to acquire knowledge". Is to say, Drescher defend the posture constructivista (not all the knowledge is to priority), but indicate the need of the existence of a knowledge congenital. When the "knowledge innate" to the that it is refer to the posture nativista, is a goal-knowledge (knowledge on how to acquire knowledge), both postures it is compare.
For other part, many authors coincide in to distinguish two types of learning:
- The responsible of to extract knowledge of the environment and to carry it to the mind of the agent.
- The learning produced through the combination and utilization of the knowledge that already it is there is acquired, that increase the efficiency of said knowledge.
In psychology it is tend to define learning as a name collective that designate "various processes complex that lead to the change of the conduct to slant of the experience". In the case that us occupy it is will make reference to "to learn" as the fact of to acquire knowledge that previously not it is were possessing. This idea of learning it is consider vulgar and wrong in psychology had to to that it is forget of the process global that suppose the change of conduct. Without embargo, it will be more useful and simple in a principle, to center the topic from this point of sight, for to include after the other aspects.
After of these positions already it is puede to clarify in certain measure the meaning of "learning without knowledge previous".
The first conclusion to the that it is arrive continuing the reflections of Drescher is that not puede to exist "learning without knowledge previous pure", already that always is necessary a knowledge previous for that it is could to give learning, that it will be the knowledge on how to acquire knowledge.
This explanation is acceptable from a point of sight reduccionista, but such time not it would be from the vision conexionista, in the that such time would result more adapted to assert that exist of form congenital a structure neurofisiológica global that permit to acquire knowledge, but not a knowledge on how to acquire knowledge, located in a level monopoly.
From the point of sight reduccionista, in a first approximation the objective of the study of the learning without knowledge previous will consist in to define the knowledge minimal that it must to possess to priority a system for that it is could to produce learning.
Exist other aspect evident, and consist in the fact of that a time that they have acquired certain knowledge, it is it will have to to attempt to employ these for to solve problems, already they will be:
-problems of the environment in the that it is inmerso the subject, is to say, the problem of to obtain new knowledge to to depart of the that already it is possess with the intention of to arrive to the knowledge necessary for to take a decision correct.
-the problem of how to optimize the methods of acquisition of knowledge, or how to create other new.
For other part, furthermore of to guide the definition of this knowledge minimal to the construction of a system adaptativo, capable of to solve problems radically different between yes, would be desirable that said knowledge minimal it would be sufficient for to produce the learning of any knowledge that could be acquired and that it would be necessary for to obtain the objective.
From this point of sight, the interest it is would center in to seek the knowledge minimal that it must to possess to priority a system cognitivo for power to accomplish to to depart of he, processes that obtain the maximum learning in the minimal time possible, or it that is it same, to to depart of small quantities of knowledge. And additionally, it is would intend that it would be sufficient for to produce, in the smaller time possible, the processes cognitivos superior that certify the intelligence.
[ Back to Index ]
The theories Psychological
Great quantity of studies on learning it is bases in the results of the investigations on how it is accomplish the processes cognitivos in the beings live. In this line of investigation they have worked numerous authors between the that fit to emphasize Pavlov, Piaget, Rescorla and Wagner, Schlimmer and Granger, and Dickinson and Shanks.
It is propose different models of establishment of relationship causative between two events -assuming that all relationship causative has place in presence of multiple causes potential-, in function of the force asociativa assigned to each event. Many authors emphasize the influence of the motivation in the learning, indicating that the learning it is produces with greater intensity when the facts affect to the attainment of the own objective, that when not it make.
Thus, the "model of rules of Rescorla" is amplified for the theories asociativas of Pearce, Schlimmer and Granger, Gluck and Bower, and for the "model asociativo of Rescorla and Wagner". The theories asociativas criticize the proposals classic of representation bases in rules, for to limit its action to concepts well defined, and propose a structure gradual of concepts, implemented normally through a mechanism of assignment of weights [Mat93]. The models asociativos avoid the rules cause-effect as such, and communicate mostly with the models conexionistas proposing a multiasociación gradual. The experiment of the approach of the theories asociativas in human it is increasing considerably in the last years.
In how much training of concepts, exist assorted opinions; from the models of prototypes of categories bases in the storage of a case ideal representative of the trend central of the examples of the category, until the models of exemplary, that deny the abstraction and defend the theory of that not exist concepts, but solely instances individual.
Exist many other aspects investigated in the psychology of the learning. For example, Neisser and Ween proven that it is learn before the rules of affirmation and denial, resulting more difficult of to learn the conditional and bicondicional [Mat93].
Other topic discussed is the of the existence of condicionamiento toward back. In the example classic of to attempt to obtain a Response Conditioned RC (salivation) before a Stimulus Conditioned EC (sound), associate to a Stimulus Incondicionado (food), the condicionamiento toward forward would suppose the presentation of the stimuli in the following order: first EC (sound), and to continuation EI (food), for to produce thereinafter and after of a certain number of trials, RC (salivation) solely before EC (sound). In a condicionamiento toward back it is would try of to observe RC (salivation) before EC (sound) when in the training the EI (food) beginning and end before of to present the EC (sound).
[ Back to Index ]
The majority of the systems bases in the knowledge they are objetivistas in how much to that contain information pre-categorized for the designer of the system (properties, objects, rules, goal-knowledge, etc.).
State of the art in the models computacionales of learning
In opposition to the objetivismo exist the subjetivismo of other authors that intend that the own agent intelligent it will be capable of to create its own categorizations in function of the signs received for its sensors. The paradigm objetivista center its attention in the levels high of the cognition, while that the subjetivista would correspond to levels low [Rod94].
Exist very few systems computacionales subjetivistas, and for it general it is try of systems created with the end of to experience aspects concrete without to arrive to it be proposed a model general acceptable.
A system of learning without knowledge previous, furthermore of be subjetivista, it must be capable of to learn in situations totally new. The mechanism for the acquisition inductive of knowledge (to to depart of the experience) it will be base in the observation passive of the environment and in the feedback (perception) of the effects of the actions accomplished, well because it is do not know to priority those effects, well because though it is they can to anticipate with certain probability, not it is has the total certainty of that the events it is develop of determined form. It is try of the learning for trial and mistake (trial and mistake learning).
Some authors indicate the learning to to depart of examples as other form of learning inductive, distinguishing three types in function of the source that provide the examples [Kod88]:
- A "teacher"
- The environment
- The own system
With the purpose of to annotate the work, it is will dispense in a principle of these possibilities. Of any form, the first of the types it is discard for to suppose a certain pre-categorization of the knowledge, that it is there is of to avoid in it possible in the search of learning without knowledge previous.
In how much to the second type, suppose a case particular of forecast, that is precisely of it that it is it is trying. In this case, it is has in account that the environment not only provide a series of situations or stages that it is they have of to analyze, and with the which it is will establish relationships cause-effect, but that the environment same provide of some form the rules cause-effect that it define, to the less in part, as it would make a book of text or a system expert. In this case, it is try of to establish relationships between the various rules cause-effect, for to predict other rules to to depart of these.
The third it is puede to assimilate to a form of representation of the knowledge. Consist in that the system, to to depart of the cases analyzed, to to depart of the rules that possess, believes other rules, of form similar to the case previous.
The utilization of these two types of learning it is it will be able to study in phases more advanced of the project.
The systems of acquisition inductive of knowledge work with a joint hierarchic of concepts, actions and rules (to the that it is add weights in a model asociativo), guided to the attainment of a goal (Goal-Oriented Behavior). The concepts they are created of form rising or Bottom Up to to depart of the experiment.
This experiment it is produces in a first moment totally to the random. After of a action it is observe its consequences and it is produces a association mental cause-effect that it is store in a report, it being indicated of some form if its execution there is favored or hindered the attainment of the goal, and in that degree. The subsequent actions they will be conditioned for the knowledge stored in the report, it being avoided in it possible to appeal to the random.
Evidently, the goals or objective of the entity intelligent they have of be defined to priority, and though thereinafter they can evolucionar for yes same, this evolution it is will give more in the form (submetas) that in the fund (goal or goals last), always that not exist a modification expresses (that it would transform in other entity different). The definition of the goals puede be complicated, in function of the behavior that it is wish to produce; and the analogies with the be human they are source of multitude of discussions philosophical on the determinismo (¿The man it is determined for a objective that it must to fulfil?), the existence physical of the soul (¿It is puede to build a be identical to the man?), etc.
[ Back to Index ]
How to learn
Drescher describe a mechanism of learning without knowledge previous guided to the computation, call schema mechanism [Dre91]. The procedure consist in to identify each possible state, situation or context in the that it is puede to find the entity intelligent through a joint of values of some items or variable that define that situation.
Exist a series of actions, primitive or operating whose execution or application makes evolucionar to the system of a state to other.
If it is accomplish a action and it is passes to a state done not know, it is store information that represent that new state and it is create a rule (schema) that represent the conditions necessary for to happen to that new state.
The problem consist how to identify what action there is been the responsible of a transition. It is recall that the actions and the states it is organize in a structure hierarchic. Is evident that if it is analyze with the same intensity (devoting the same time) all the possible connections action-transition in all the levels of the hierarchy, the time of process would tend quickly toward infinite. Here enter in game the theories psychological commented previously.
Schlimmer and Granger developed a algorithm call STAGGER that was implementing the theories asociativas through the assignment of weights. STAGGER is capable of to detect noise or information imperfect, accomplishing a processing provisional of the data, always that the noise that it is produces it will be of type systematical -only in the examples, or only in the contraejemplos-. IKASLE is a algorithm that operate of way similar to STAGGER, being capable of to learn in cases of noise random [Alb92].
The project SOAR it is appointment for be one of the more underlined efforts in the definition and creation of Architectures for the Intelligence General, already that include many of the characteristic of other systems that it is they have built with precedence to he, and because its architecture is checked continually adding new possibilities to the new versions of SOAR.
The Architectures General of Intelligence bases its operation in a joint hierarchic of submetas. The submetas they are created as effect of the attempt of to arrive to other goal of form falling or Top Down. Without embargo the system, to priority, not puede to decompose a goal in a submeta, already that not possess the knowledge necessary for to accomplish said decomposition. For this reason, the architectures bases in the creation of submetas, not provide a mechanism sufficiently potent for to produce learning without knowledge previous.
After of have checked the different theories and proposals applicable to the creation of a system cognitivo artificial capable of to accomplish learning without knowledge previous, it is observe that for a part exist diversity of opinions in how much to the existence or not of different aspects in the learning (case of the models of exemplary that deny the existence of concepts; case of the condicionamiento toward back).
For other part happen that for it general exist more of a procedure for to carry to the practical a same theory; for example, a law that define the effect of the motivation for the assignment of weights in the creation of relationships causative will possess constant that there is that parametrizar of some form.
Other example that puede be very illustrative -already that it is try of one of the more simple-, is the of the spread supposed of the existence of various reports (of receipt sensorial, to i cut term, and to long term) for the retention immediate, to i cut and long term, as parts of the process of learning.
Is evident that the capacity of report is innate. If it is wish to include these aspects of the report in a system of learning not to priority, emerge multitude of questions: ¿It is specify two or three types of report? ¿Or yet better it is make possible the existence of a number variable of types? In that case, ¿Variable in function of what? ¿It is permitted the elimination of zones of report? ¿Under what conditions it is permit the erased of a datum? ¿How it is divide the capacity total between the different reports?
[ Back to Index ]
Life artificial and systems emerging
The behavior emerging consist in the phenomenon of it be produced a behavior complex to to depart of a series of rules simple.
A example illustrative of the behavior emerging is the case of the experiment developed for Cristopher Langton in the University of Michigan. Langton simulated a environment artificial of a colony of ants in a grid compound for celdillas blancas, blue and yellow. The rules assigned to the ants they were: if a ant penetrate in a celdilla white, continues it being displaced toward forward. If it is introduce in a celdilla blue, this adopt the color yellow, and the ant will turn to the left; and if penetrate in a celdilla yellow, this it is will become blue, and the ant will turn to the straight.
In a first moment, the ants erred without course defined, but in little time the ants it is found some to other forming a string. The aspect interesting of this experiment is that not it is there is programmed explicitly that the ants adopted a behavior social of this type, but that it is produced spontaneously. This type of behavior is customary in experiments of life artificial.
Note: this experiment it is there is intentado to repeat to the less for three persons that conozco without to achieve no success. All the references to the experiment original of that dispongo they are indirectas. All seem to indicate that the experiment such as it is there is described it is incomplete or wrong outlined.
Some of the doubts that emerge they are:
- If a ant it must it be located in a cabin in the that it is find other ant, ¿what it must to make? ¿it be located on this? ¿to wait to the following iteration? ¿to change of course aleatoriamente? And related with this: ¿The state of all the ants it must to change "to the time", in cycles, in function of the states previous of all the cells, or we can to execute the ants a for a in sequence?
- ¿Is sufficient with those rules for that surja the such "behavior emerging", or is necessary some other?
- ¿Of what dimensions there is of be the grid?
- ¿How many ants is necessary to put?
- ¿What address initial it must to carry each ant? ¿Random?.
- ¿What percentage of cells they should be blancas, blue and yellow?
- ¿It is they should situar the cabins blue and yellow of form random, or forming some standard special?
- If a ant arrive to the limit of the grid continuing its address, ¿what of it must to make? And related with this ¿They can or they should to exist celdillas prohibidas that not they will be neither blue neither yellow neither blancas?
Many authors coincide in to indicate that the learning of the organizations has place, furthermore of in the organization individual, to level of kind. The learning of the kinds it is manifesto in the evolution, it will be formed for characteristic present in all the individual of the kind, and puede to originate behaviors emerging as the mentioned. Puede it be imagined also a learning to level of ecosistema, in the that intervene several kinds. The evolution biological it is base in two processes: the selection and the variation genetic. The genes better endowed they will be the prevailing. The process not experience stagnation because past certain time it is produce mutations, and appear new genes and new combinations of genes through reproduction sexual.
Thomas Ray designed in the University of Delaware a surprising system call LAND capable of to imitate in great measure the processes evolutionary. In he, each organization is a program with capacity of autorreproducirse, it being found all they in a enclosure of report limited. When a program find a zone of report free, copy its own code. It is permitted modifications to the random to the hour of to retort a program; it is included a program that was eliminating to the programs more old or those whose code was wrong, it is established that the execution of certain instructions could to delay the death, etc. In a principle it is introduced a only program of 80 instructions. To the end of a time emerged a program of only 22 instructions capable of autorreproducirse to a speed much greater that the of the creature original; appeared parasite unable of it be reproduced, that were robbing code of the programs adjacent. These parasite they were surpassed for some programs exempt to they, until that appeared other better endowed, unable of it be reproduced in isolation but capable of to make it in community.
To continuation it is is going to to develop the hypothesis of that the processes cognitivos in a alone entity intelligent, as puede be it a person, they can be formed for the result of the interaction of multiple entities simple endowed each a of they of the characteristic own of the beings live, it being observed from the foreign that emerge a result coherent fruit of a situation interior complex.
Independently of the validity of this hypothesis, a mechanism of this type would serve of base for the implementation of the theories asociativas offering a mechanism much more complex and potent that the simple assignment of weights and would serve to its time as laboratory for the experiment of the validity of the various theories psychological.
The position of Drescher in the discussion between the postures constructivista and nativista before the problem of the existence of knowledge innate it is puede to apply on yes same so many times as it is wish in a varying of the paradox of Russell on the car-observation, similar to the hypothesis of Rucker on the mind humanizes as a system infinitely introspective, and in the line of many of the reflections of Hofstadter on the theorem of Gödel [Mor91] [Hof87].
We call "To" to the knowledge on how to acquire knowledge. The knowledge on how to acquire knowledge (knowledge "To") would not be innate, always that existed, of form congenital, a knowledge "B" on how to acquire knowledge "To", is to say, a knowledge on how to acquire knowledge on how to acquire knowledge. The knowledge "C" would be the knowledge on how to acquire knowledge "B" and thus sucesivamente.
A system with knowledge "To" and "B" it will be more potent and more general that one that only possess knowledge "To", and a system with knowledge "To", "B" and "C" would be more general that one with knowledge "To" and "B".
It is go to to describe with greater detail each one of these levels. If it is intend to build a system artificial that contains levels of this type, the characteristic principal that they should to fulfil these levels is be recursivos. The recursividad it is understand in how much to that the mechanism that provide its existence it must be only.
Each one of the levels consist in a resolutor of problems, already that it is puede to identify the acquisition of knowledge as a problem to to solve. The first of they analyze a problem whose elements belong to the world real, formed for the entity intelligent and its environment. The second analyze the behavior of the first, the third the of the second and thus sucesivamente.
Defining instance as representation mental of a object or occurrence specific, and class as representation mental of a group of objects, we have:
- In the first level the entity intelligent analyze the world foreign to he and accomplish the following processes:
- To create instances of level 1 that represent situations or states of the world foreign. These they will be the only that will will have its equivalent in the world foreign. Each instance represent a possible state of the world foreign.
For to create instances it is will use methods of creation of instances, that will consist in the simple observation of the "world foreign".
- To create concepts to to depart of the generalization of instances. These concepts only have existence in the world interior of the thought.
For to create concepts it is will use methods of creation of concepts bases in to compare and to merge instances.
- To solve problems of the "world foreign" handling concepts and instances.
For to solve problems it is will use methods of resolution of problems bases in the feedback of the effects of the actions executed in the "world foreign"
Exist a series of parameters that define the operation of this first level, of way that each combination of those parameters configure a method of resolution of problems different.
- In the second level, it that it is analyze not is the "world foreign" to the thought but the own thought. It is analyze the resolutor of problems of the level 1. the processes they are:
- To create instances of level 2 that represent situations or contexts of the level 1, is to say, methods of resolution of problems used for the level 1. Each instance represent a state of the level 1. The methods of creation of instances, will consist in the observation of the first level.
- To create concepts to to depart of the operation of generalization (to compare and to merge) applied to a joint of instances that represent methods of resolution of problems.
- To solve problems of the level 2 handling concepts and instances. In this case, the "actions" correspond with the election of the parameters with the that operate the first level. Thus it is will obtain new methods of methods of resolution of problems (goal-methods). From the point of sight of the knowledge, it is will acquire goal-knowledge.
The operation of this second level it is also defined for a series of parameters, that they can be handled for the level 3, and thus sucesivamente.
It is observe the symmetry between the processes developed in each one of the levels, that will permit its implementation recursiva. Also it is makes to note that the level more high of the hierarchy not is controlled for no other, already that not exist a level superior to he that it observe, and its existence define the limitation of the flexibility of the system.
We imagine now the existence of a super-level that possess knowledge about of how many levels of knowledge they are desirable, of way that this level could to decide the number of levels from "To" until "Z" that they should to exist. This super-level would to judge even its own reason of be; would to decide that its existence not it is justified already that the number of levels optimum is, for example, three; and to decide it be eliminated to yes same.
So much in the man as in a system cognitivo artificial, is possible that exist knowledge of the type "To", "B", "C", etc. Currently not exist studies psychological that determine if the man is a system intelligent with a number fixed of levels of this type or if possess a super-level that to him permit to possess a number anyone of they, though it is puede to intuit the existence of to the less three or four levels.
A system artificial it will be able to possess so many levels of this type as we be capable of to provide to him. I will call "Z" to the last level of these "knowledge". Though the number of levels not it will be able be infinite, a Resolutor General of Problems or GPS (General Problem Solver), it will have to to possess the capacity of to have of any number of these levels of knowledge, always that the existence of said levels it will be indispensable for to solve a problem.
We can to suppose that never it will be necessary to use more of a certain number of levels. In that case, the creation of this system will serve for to discover which is the level "Z" to the that it is would have to to arrive in the construction of a Resolutor General of Problems. For the opposite, if it is proven that in some cases is of great usefulness the existence of a great number of levels, and it is they could to identify the types of problems that require a or other structure of levels, it is would give a great step in the construction of GPS.
In this position, the concept of knowledge it is returns confusing. Is customary to divide the knowledge in declaratory, procedural and heuristic. A fact concrete it is would represent of form declaratory. The mechanism for to obtain new knowledge would be a knowledge procedural. The knowledge heuristic is a knowledge complex that in the better of the cases we can to suppose it as combination of the two previous (in the worse of the cases, it is does not know how it is represent this knowledge).
We suppose that is possible to accede to the symbols physical that represent a goal-knowledge given, is to say, a knowledge of type "To", "B", etc. or super-knowledge, and to modify said knowledge to the random, it being understood the random as the election of a possible action (modification) between several in function of a cause not controlled or unknown.
This modification to the "random" would be a mechanism capable of to generate new goal-knowledge. The usefulness of this new knowledge would be doubtful in a principle, already that the probability of that the new knowledge generated operate as a mechanism of learning valid would be normally small.
Without embargo, in the case of the learning without knowledge previous in the "world foreign" (in the case of the acquisition of "knowledge" to you dry), the mechanism it is base initially in the execution of actions to the random. For analogy, the learning of knowledge superior (type "To", "B", etc.) in the "world interior" would it be produced of the same form.
Exist the hypothesis of that the origin of the intelligence in the various kinds animal and in the man consist in the combination of elements to the random, in mutations genetic and in the mechanisms evolutionary of the selection of the kinds -that let a door opened to positions philosophical, religious, etc. in how much to that not intend to answer to the question of what or who created the chaos initial-.
continuing this hypothesis of a form analogous, the knowledge minimal necessary for the creation of a system capable of to accomplish the processes cognitivos superior would be a knowledge stored of form procedural that consist basically in the random in each one of the levels and goal-levels mentioned previously.
It is recall that this random only it is produces in a state initial and that to to depart of that moment it is try of a random directed for the goals. The mechanism of struggle for the survival it is would entrust of seleccionar in a first place the entities whose processes of random you they have carried to the attainment of the objective -it is puede to think in the goals in terms of the pyramid of needs of Maslow (in first place the needs basic physiological, etc.)-. Thereinafter, in a environment adverse of shortage, it is seleccionarían the entities capable of to represent, to store and to use the experience acquired in the attainment of objective for the satisfaction of new needs.
I propose that the support that permit the creation automatic of these processes to to depart of a knowledge minimal in a system computer it will be a system of Artificial Life, in the that, so much the instances as the classes, the rules and the mechanisms of learning and all the elements that it compose they will be objects of entity own that include each one of they the characteristic own of a be live: execution parallel, objective, birth, death (in the case of not to obtain the objective), reproduction, mutations genetic (random), etc.
This support would be capable of to simulate the existence of several entities intelligent, to include aspects as the selection of the kinds, the collaboration between individual, etc.
Also would be the support suitable for to include the rest of the aspects that certify the learning and that it is they had omitted in a beginning: motivation, fear, desire, usage, perception, etc.
[ Back to Index ]
KLEIN, STEPHEN B.
1994 APRENDIZAJE. PRINCIPIOS Y APLICACIONES
1990 THE STRUCTURE AND FORMATION OF CATEGORIES
en THE PSYCHOLOGY OF LEARNING AND MOTIVATION
Bower, G. H. (Ed.)
PEARCE, J. M.
1980 A MODEL FOR PAVLOVIAN CONDITIONING
RESCORLA, R. A.
WAGNER, A. R.
1972 A THEORY OF PAVLOVIAN CONDITIONING
en CLASSICAL CONDITIONING II
Black, A. H. (Ed.)
1989 FOUNDATIONS OF COGNITIVE SCIENCIE
1993 APRENDIZAJE Y REPRESENTACION DE CONCEPTOS
en APRENDIZAJE Y MEMORIA HUMANA: ASPECTOS BASICOS Y EVOLUTIVOS
J.I. Navarro (Ed)
1992 APRENDIZAJE A PARTIR DE EJEMPLOS Y ASIGNACION DE PESOS
en NUEVAS TENDENCIAS EN INTELIGENCIA ARTIFICIAL
Moral, A. y Llaguno, M. (Ed.)
Universidad de Deusto
ROSENBLOOM, PAUL S.
1993 SOAR PAPERS: RESEARCH ON INTEGRATED INTELLIGENCE
Volúmenes 1 y 2
ROSENBLOOM, PAUL S.
1986 UNIVERSAL SUBGOALING AND CHUNKING
Kluwer Academic Publishers
1993 TOWARDS A UNIFIED MODEL OF LEARNING TO PROGRAM
en COGNITIVE MODELS AND INTELLIGENT ENVIRONMENTS FOR LEARNING PROGRAMMING
Lemut, Enrica (Ed.)
NATO ASI Series. Series F, Computer and Systems Sciences
SIMON, HERBERT A.
1972 HUMAN PROBLEM SOLVING
1993 THE INTERACTION BETWEEN KNOWLEDGE AND PRACTICE IN THE ACQUISITION OF COGNITIVE SKILLS
en FOUNDATIONS OF KNOWLEDGE ACQUISITION
Susan Chipman (Ed.)
Kluwer Academic Publishers
BEER, RANDALL D.
1990 INTELLIGENCE AS ADAPTIVE BEHAVIOR
ANDERSON, JOHN R.
1990 ADAPTIVE CHARACTER OF THOUGHT
Lawrence Erlbaum Associates
STADDON, J. E. R.
1983 ADAPTIVE BEHAVIOR AND LEARNING
Cambridge University Press
1988 INTRODUCTION TO MACHINE LEARNING
Morgan Kaufmann Publishers
HINTON, GEOFFEY E.
1989 PARALLEL MODELS OF ASSOCIATIVE MEMORY
Lawrence Erlbaum Associates
KLAHR, DAVID (Ed.)
1987 PRODUCTION SYSTEM MODELS OF LEARNING AND DEVELOPMENT
DRESCHER, GARY L.
1991 MADE UP MINDS
HOFSTADTER, DOUGLAS R.
1987 Gödel, ESCHER, BACH. UN ETERNO Y GRACIL BUCLE
Tusquets Editores & CONACYT
MORRIS, HAROLD C.
1991 ON THE FEASIBILITY OF COMPUTATIONAL ARTIFICIAL LIFE
en FROM ANIMALS TO ANIMATS
Meyer, Jean-Arcady y Wilson, Stewart W. (Ed.)
LANGTON, CHRISTOPHER G. (Ed.)
1989 ARTIFICIAL LIFE
The proceedings of an interdisciplinary workshop on the synthesis and simulation of living systems
1990 ARTIFICIAL LIFE II
Proceedings of the workshop on artificial life
MULLER, JEAN PIERRE (Ed.)
1994 VIRTUAL ENVIRONMENTS FOR SIMULATING ARTIFICIAL AUTONOMY
en ARTIFICIAL LIFE AND VIRTUAL REALITY
Magnenat Thalmann, Nadia (Ed.)
John Wiley & Sons
¡You are guest to hold, criticize and collaborate with your ideas!
Send your commentaries, questions, suggestions, or criticize to to E-mail
[ Home Page Castellano | Home Page English ]