A conceptual framework, whose goal is the improvement of efficiency of machine learning, is presented. The framework is designed in a broader context of problem solver (PS). The design is solved as an integration of all basic cognitive functions and as a software-engineering problem. Many (one hundred) requirements imposed on PS are considered. The most important of them are the object-oriented nature of the PS environment, reflexivity of PS, and central role of tool and shifted border.1
This is a position paper, presenting a framework for machine
learning (ML), reacting on some reviewers’ reamrks. The
framework is very briefly described and justified. More details can
be found in [
My extensive evaluation of current ML reveals that the
outstandingly dominant part of ML is devoted to simple learning. I
consider this harmful and the framework is a trail to remedy this
situation. I would roughly characterize simple learning as a
oneshot creation of knowledge, describing one function of a small part
of an environment, and applied in a manually selected part of
environment. On the other hand, I would characterize complex
learning as a cumulative creation of knowledge with many parts,
which describes many related functions of whole environment. It is
structured, preferably in object-oriented way, applied in different
parts of different cognitive functions, applied in the whole real
world. These are two extremes. The state-of-the-art is, of course,
somewhere in-between, much closer to simple learning. It contrasts
with some other AI areas, like knowledge engineering, where we
do use complex knowledge structures [
The application of this framework to learning ontologies is
highly relevant: 1) Both in ontologies and in the framework, there
is a common stress put on knowledge structure, reuse, object
orientation. 2) Learning ontologies can be approached both top-down
and bottom-up. If there are no worked out firm foundations then
they should be done first. This is also the aim of the framework. 3)
Both learning and ontology areas can be integrated. For example,
learning could accept complex knowledge structures; ontologies
could accept an approach to approximate knowledge. This would
modify e.g. some design “principles” that “have been proved useful
in the development of ontologies”, like clarity, completeness, and
coherence [
My approach to learning can be characterized concisely as "learn from learning". This means that it is suitable to use knowledge, which we have gathered during the exploration of learning (phenomenon), and apply it as meta-knowledge in the design of improved learning (algorithm). The body of this knowledge is vast. What parts should be used? My approach is to use all meta-knowledge that can be integrated into an efficient learning system. It is also the solution of the efficiency of learning: Each piece of this meta-knowledge should support somehow this efficiency. To design and build a corresponding system is a formidable task. However, realize four things: 1) There is a great redundancy in AI knowledge. 2) There is a difference between the building of a learning system and fully developed problem solving system. I am concentrating on problem solving (PS), which has not yet gathered a massive amount of knowledge, but which can gather it. I am using the frame of PS to stress the importance of the context of learning. 3) There are approaches to the design of complex systems; I am using the software-engineering approach and the object-oriented methodology. 4) There is no evidence that this task is unsolvable.
The main objection against the proposed framework is why should we deal with something what is so unjustified, unproven, unimplemented? Haven’t we done it already many times without any result? Is not here a rule that every proposal should be at least partially verified by some prototype? I understand this but do not agree. Why? 1) Critical claims are usually very vague. 2) Even negative experience is useful, it exists, and I’m trying to use it. 3) Prototype verification has its value; however, it should not be overrated. 4) Not only a detailed analysis, the synthesis can bring new knowledge too. To prototype a synthesized system is harder and needs more careful design. 5) It is in contradiction with the experience of software engineering: design should be verified step-bystep. 6) AI should realize the shortage of synthesis, its reasons and consequences, and should try to solve it.
In the design, I primarily use the viewpoint of efficiency of
learning [
First, I am presenting basic assumptions about the learning context. They form a skeleton for the requirements: • Environment is a network of many heterogeneous, dynamic and even uncertain objects. • PS is an object of environment. PS has a pre-specified goal. If PS is not in goal state, it has problem, and that should be solved by PS. PS should control environment to reach goal. PS control is accomplished by means of its cognitive functions (implementation, identification, reasoning, learning, knowledge base (KB), self-control, and initialization).
The cognitive functions have also their (sub) goals. • Knowledge is an approximate description (of the behavior) of object; knowledge usage increases the average probability of success of reaching PS goal. • Environment is relatively very stable. During an interval of PS work, very little part of environment is changing. • PS (learning) should cope with its complexity.
The framework itself is outlined in two figures. To fulfill the
requirements, PS should be a knowledge-based system and KB
should have the structure outlined in Figure 1. (For graphical
notation see [
Learning is not a stand-alone function. It cannot work without cooperation with other PS cognitive functions. PS should have the structure outlined in Figure 2. Notice that it is still a partial view, e.g. relations to important self-control and initialization functions are not rendered here. Notice also, there is only one common concept, description, for description of various PS entities, i.e. classes of environment objects, specific objects like model, plan, value, variable, state, goal, meta-model, meta-plan, etc. 3
Here I show two examples of requirements, tool and shifted border, ,GHQWLILFDWLRQ 5HODWLRQ W\SHV ,QVWDQWLDWLRQ 6LPSOLILFDWLRQ $JJUHJDWLRQ $VVRFLDWLRQ (TXLYDOHQFH $QFKRU ,V JRDO RI DWWULEXWH
LQVWDQWLDWLRQ 5HODWLRQ DWWULEXWH
’HVFULSWLRQ
VXESDUW LQVWDQWLDWLRQ
LQSXW
RXWSXW
VWDWH ,QWHUIDFH
UXOH
5XOH
and homogeneity, and briefly explain how they are implemented in
the framework. More can be found in [
In a simple environment, to identify (the only one and simple)
object and its state need not be a problem. To learn description of
such object, plan a solution in this environment described as such
object, and implement plan need not be a problem. Self-control
would not be necessary and KB would be trivial. To implement it,
the existing cognitive functions or ontologies from AI can be used
[
Some aspects of object-oriented nature of environment are solved by existing techniques, 0RGHO e.g. instantiation, simplification, and anchor relations in learning. To solve other aspects, e.g.
association relations, the con5HDVRQLQJ cept of tool is used: Let us consider PS, its environment, some tool, as an object of environment, and rest of environment. Let PS have inputs resp. 3ODQ outputs to perceive resp. influence environment; the part of them connected with tool is x resp. y. x and y together form the part of the real border be,QSXW 2XWSXW 2XWSXW (QYLURQPHQW tween PS and environment, i.e. border between PS and tool. x' and y' form the part of border between tool and rest of environment. x' and y' are accessible to PS during learning but not fully accessible to PS later. Let us assume that PS can learn characteristics of tool in such a way that it can later 1) identify tool, 2) model (specify) the sequence of values of x' and y', if it knows the sequence of values of x and y. x' and y' are called modeled inputs and outputs. The equality of modeled inputs and outputs, both in environment and in PS, can be interpreted as a shift of border between environment and PS. We can interpret e.g. teacher, society, communication with similar PS's, and an approach to learning as tools.
There are approaches using e.g. genetic algorithms for
self-control [
Let us consider the possibility to implement, in the framework, two types of descriptions, i.e. description of object of environment and description of PS plan. These two descriptions should have had many similar properties, e.g. they should have been general, dynamic, complex, hierarchically organized, object-oriented, might have been approximate, uncertain, can utilize tools etc. How might they have differentiated? The only difference is that (sub-)plan should be directly related to some (sub-)goal. There is also another viewpoint: PS and its parts have their behaviors. They are described - generated by plans. PS and its parts are parts of environment. For both, it is suitable to have common description. As a description of external object and plan share the majority of features, it is suitable to describe them in a common way. 4
This paper proposes a framework to improve the current level of efficiency of learning. The approach shows feasibility in a sense that, on the conceptual level, the PS could be designed to satisfy all challenging requirements. The efficiency of learning has been analyzed and designed on a conceptual level. This is the only possible (and necessary) beginning of a long-term project. My intention is to continue with a detailed analysis and design of cognitive functions and KB (knowledge base), and their implementation using a contemporary software-engineering approach, i.e. utilizing object-oriented, computer-aided software-engineering approaches and tools and using contemporary knowledge about cognitive functions.
The idea to integrate all AI concepts into one system is not new. However, integration is often preached, but seldom practiced. My contribution is in doing the "first (small) step" toward such integration. The framework uses many concepts, but not many new ones. The exceptions are “learn from learning”, shifted border, that learning can create all parts of KB, and knowledge homogenization.
In higher cognitive processes, self-reflection plays an important role. AI is a cognitive process of incremental understanding the phenomenon of intelligence. What is the state of self-reflection in AI? According to my assessment, the field of AI is just at the beginning of its conscious self-reflection. What do I mean by this? I mean that parts of AI community - yet still small - are just starting to use specific benefits of AI work, i.e. they are starting to interpret the results of the AI process - the knowledge produced by AI - to control the AI process itself. I have tried to use this knowledge already in this paper.
7KLV ZRUN KDV EHHQ SDUWLDOO\ VXSSRUWHG E\ WKH JUDQW *$ 5 1R 102/99/1564. I also acknowledge the help of the reviewers, my colleagues J. Grim and A. Halousková, M. Kárný, my daughters, and niece.