Survey
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project
Software Engineering of Granular Systems Granular Computing and Hypercomputation Andrzej Bargiela NAFIPS 2004, Summer School Overview and motivation Granular Computing came about through the realisation of incompatibility between human and machine information processing Fuzzy and rough set formalisms of information granulation provided an excellent framework for knowledge representation (Zadeh, Pawlak, Kacprzyk, Yager, Kandel, Bezdek, Pedrycz, Kreinovich, Hirota, Skowron, Slowinski) The key remaining question is whether the processing of information granules (knowledge) can be represented adequately as a computational process. Workshop Plan Fundamentals of computation Information vs. knowledge processing Hypercomputational processing of knowledge as an aspect of human intelligence Information granulation fundamentals Granular Computing Granular Computing and Hypercomputation Fundamentals of computation Definitions Conceptual “computation” Formal behaviour equivalent to UTM (so “computable” means “computable by UTM”) Physical “computation” Action of a physical system that implements conceptual computation (man or machine) Implementation that implements computable function is a “computational system” Universal Turing Machine UTM Finite but unbounded tape for storage of symbols A head for reading and writing symbols from/to the tape Finite set of symbols used for input/output. Finite set of internal states Transitions associated with internal states and detailing how the machine should function (symbol1, symbol2, direction, new-state) Universal Turing Machine UTM to calculate binary expansion of 1/3 Initial tape Final tape 0 0 0 0 0 0 0 0 0 0 … 0 1 0 1 0 1 0 1 0 1… 0/0/R/1 1 0 0/1/R/0 Universal Turing Machine Turing’s halting function Uncomputable set Function that takes as its input a natural number representing a Turing machine and returns 1 if the Turing machine halts or 0 if it does not. {n | n represents a Turing machine / input pair that halts} Uncomputable real x such that its n-th digit of its binary expansion is 1 if n represents a Turing machine / input pair that halts and 0 otherwise Universal Turing Machine Proof of uncomputability of a halting function Assume there exists a Turing machine T for the halting function H Construct a modified machine T* which takes an encoding of T and determines whether T halts when given its own encoding as input Construct another machine T** that is like T* but it loops if T halts on its own encoding and halts if T loops on its encoding If T** is given an encoding of itself as input, it loops if and only if it would halt. A contradiction T does not exist Universal Turing Machine So, what can be evaluated by UTM? Church-Turing [CT] thesis “… the numerical functions that can be effectively evaluated by human clerical labour, working to fixed rule, and without understanding are precisely functions that can be evaluated by the UTM …” Interpretation of CT thesis Misunderstanding ! Any function that can be computed by any means whatsoever can be computed by UTM Correct interpretation Humans can in-principle, simulate computers through clerk-like behaviour The computer behaviours are precisely the effective human behaviours but … human behaviours go beyond that!!! especially when it comes to knowledge processing Illustration of CT thesis Intelligent human behaviours effective computer = (clerk-like) behaviours behaviours Effective = clerical labour, working to fixed rule, and without understanding processing of symbols (information) Characteristics of behaviours Effective Processing of finite set of symbols (information) Operation to “fixed rule” Operation without insight Operation without understanding Intelligent Processing of uncountable abstractions (knowledge) Keeping “open mind” about operation Essential insight Essential understanding Beyond UTM Hypercomputational system IS the Nature computational? A system that “computes” non-computable behaviours (such that cannot be simulated by UTM) All implementations are computational IS the Nature hypercomputational? There exist at least one implementable behaviour that is not computational Quo Vadis AI ? Wrong premise ! “.. If a scientific (empirically verifiable) study of intelligence (AI) is possible at all it must be capable of being expressed in computational terms..” (scientific method does not exclude insight/intelligence but definition of computability does) “.. A standard digital computer, given only the right program, a large enough memory and sufficient time can compute any rule-governed input-output function..” (unwarranted extension of the definition of UTM from manipulation of symbols to any I/O mapping) The meaning of “A” in AI AI dilemma If “artificial” in AI is to be interpreted as “fake” then it can continue with developing computational models of (ersatz) intelligence, but … If “artificial” in AI is to be interpreted as “produced by means that are not of human origin but equivalent in substance to the original thing”, AI must go beyond “computation” i.e. focus on “hypercomputation”. Illustration of CT thesis Intelligent human behaviours (that include) Hypercomputational behaviours effective computer = (clerk-like) behaviours behaviours Effective = clerical labour, working to fixed rule, and without understanding processing of symbols (information) Intelligent / hypercomputational behaviour Mathematical reasoning Goedel’s Incompleteness Theorem (1931) Principia Mathematica is either incomplete or inconsistent theory of natural numbers. Proof: Mapping of arithmetical formulae onto numbers (so that statements can refer to each other). Mapping of a statement “is provable in Principia Mathematica” onto a statement in arithmetic. Formulation of a statement that says of itself that it is not provable within Principia Mathematica. “Finite means” reasoning Incompleteness Theorem extension Any formal proof system that operates by final means is either incomplete or inconsistent Proof Mapping of axioms onto numbers (so that axioms can refer to each other) Mapping of inference rules onto statements in arithmetic Formulation of a statement that says of itself that it is not provable within a given formal proof system. “Computational” reasoning Interpretation of “finite means” in the context “mechanical computation” Turing: Finite means interpreted as a computation accomplished by a human agent operating without insight Kleene: Finite means interpreted as a recursive evaluation of functions that are rooted in a set of basic functions (zero, successor, projection, composition, primitive recursion). Same set of computable functions Entscheidungsproblem Given a set of axioms A decide whether the theorem T is or is not provable from A. Failure to find an algorithm (Turing Machine) for solving such a problem General argument: The set of functions from N to N is uncountable while a set of Turing machines is countable. Specific examples of functions that cannot be computed (Turing’s halting function) HC rationale We need hypercomputation (HC) to process knowledge as opposed to processing just an information … more than “computing” … Computing Finite set of symbols Operation to “fixed rule” Operation without insight Operation without understanding Single correct result … more … Infinite, enumerable set of symbols Infinite state Infinite time A-priori results Theoretical forms of HC Infinite Infinite Infinite Infinite Infinite decay) ! memory (Oracle, initial inscriptions) state speed time mass/energy (harnessing radioactive Infinite memory Turing’s Oracle machine (OTM) The machine is equipped with three additional special states: 0-state, 1-state and “call state” If the Oracle stores the halting set (potentially infinite) then the presentation of the input on which the machine halts can stipulate the reference to the Oracle and successful “computation” of the halting function (non harnessable hypercomputation) ! Infinite memory UTM with initial inscriptions The odd squares of the machine tape are filled with the binary expansion of the real number representing the halting set and the even squares are used for normal operation of UTM The initial inscription is effectively another form of Oracle (potentially infinite) (non harnessable hypercomputation) ! Infinite state UTM with infinite set of states The implication is that there is an infinite number of transitions from which a finite number leads to a given state The states (infinite set) of the UTM* can be used to represent the halting set of the corresponding UTM (non harnessable hypercomputation) ! Infinite speed UTM with geometrically increasing clock The first step is assumed to take 1 unit of time and the subsequent steps are taking ½ of the time required by the previous step. So the infinite number of steps can be accomplished in 2 units of time (infinite execution speed) The problem with the halting set is bypassed by taking the output from the first square after 2 time steps (non harnessable hypercomputation) ! Infinite time UTM allowed to operate infinite time The configuration of the machine is defined from the preceding configurations. The machine goes into a special limit-state and each tape square is defined as: 0 if the square settles to 0; 1 if the square settles to 1; and 1 if the square alternates infinitely often between 0 and 1 After deciding the state of a square the machine is restarted from the first square (non harnessable hypercomputation) ! Infinite mass/energy Probabilistic UTM The machine has two applicable transitions from a given state. When in such a state, the machine chooses the transition with equal probability. The machine computes a function if the probability of a correct answer is greater than ½ Machine that outputs randomly 0 or 1 with equal probability (using eg. Radioactive decay as a random process) generates a non-recursive real with probability 1 but there is probability 0 of generating any specific number (non harnessable hypercomputation) ! Practical form of HC Use ordinary computers but … use them in an “intelligent” way to do more than just “computing” … more than “computing” … Computing Finite set of symbols Operation to “fixed rule” Operation without insight Operation without understanding Single correct result … more … Infinite, enumerable set of symbols Infinite state ! Infinite time A-priori results … more than “computing” … Computing Finite set of symbols Operation to “fixed rule” Operation without insight Operation without understanding Single correct result … more … Infinite, enumerable set of symbols Infinite state ! Infinite time A-priori results Physical simulation Keep open mind about the results Hypercomputational processing of knowledge as an aspect of human intelligence … more than “computing” … 2 1 1 … more than “computing” … … more than “computing” … … more than “computing” … Practical form of HC Use ordinary computers but … allow them to be “open minded” (based on evidence derived at different levels of abstraction granulation) Information granulation fundamentals Granulation as a basis for being “open minded” Granulation - a process of instilling additional semantic content into data (not just a simple aggregation). Fundamental distinction from “flat” numerical interpretation of data. An insight rooted in the axiomatic set theory. Axiomatic set theory Georg Cantor foundations of Set Theory 1874-1884 Two primitive notions of: sets and membership Paradoxes in initial theory: - The cardinality of “set of all sets” vs. cardinality of the set of all subsets drawn from this set. - Construction of sets paradox (Russel) Axiomatic set theory Modern Set Theory (ies) - (eg. Goedel, 1940) Abandons a uniform view of sets and introduces some form of “hierarchy” Emphasises the semantical transformation (qualitative change) that occurs when grouping individual elements into sets or conversely when identifying sub-sets within a given set Three primitive notions of: class, set and membership Axiomatic set theory Modern Set Theory (ies) Class - an entity corresponding to some but not necessarily all properties (-> various classes represent conceptually different entities) A class that is a member of some other class is considered a set, otherwise it is considered a proper class (-> a “set of all sets” is a proper class) Granular Computing New Insights Granular Computing: A paradigm for computational processing of information granules A paradigm for harnessable hypercomputation through concurrent computations along the dimension of Information Abstraction New Insights Infinite Resource Hypercomputing Turing Machines Abstractions Physical Logical States States Numbers Sets Concepts New Insights Infinite Resource Hypercomputing Turing Machines Granular Computing Granular Computing Granular Computing Granular Computing Abstractions Physical Logical States States Numbers Sets Concepts Granular Computing G = <X, G, A, C> Granular world data space granulation framework communication mechanism family of granules A/D Digital processing D/A Granular Computing G = <X, G, A, C> Granular world data space granulation framework communication mechanism family of granules Digital A/D D/A Digital D/D processing D/D Digital D/D processing D/D Digital D/D processing D/D Digital D/D processing D/D processing Granular Computing Concepts: Encoded as paintings/sculptures/music etc. Original idea Decoded idea Granular Computing Concepts: Encoded as paintings/sculptures/music etc. Processing of pixels Original idea Image segmentation Texture/colour analysis Analysis of metaphores Decoded idea Granular Computing Concepts: Encoded as paintings/sculptures/music etc. Original idea Decoded idea Granular Computing Concepts: Encoded as paintings/sculptures/music etc. Processing of pixels Original idea Image segmentation Analysis of rendering Analysis of metaphores Decoded idea Granular Computing and Hypercomputation Entscheidungsproblem Given a set of axioms A decide whether the theorem T is not provable from A At the “coarse” level of granularity produce an instantaneous negative answer At the “fine” level of granularity perform a computational process of permutating axioms If the processing at the “fine” level of granularity results in proving T then change the answer to affirmative Conclusions • Granular Computing arose through insights from pattern classification • Axiomatic set theory underpinning the above insights • The prominence of fuzzy/rough sets for the description of information granules (knowledge) • Knowledge processing as an inherently hypercomputational process • Theoretical and practical hypercomputer implementations • Software Engineering of Granular Systems concurrent multi-resolution processing as a harnessable hypercomputation Granular Computing and Collaborative Computation Collaborative processing Distributed sources of data (databases) No direct interaction between databases Goal: reveal structure common to all databases through a process of collaboration Approach: explore possibilities of collaboration through granular communication Horizontal collaboration Databases about same N clients at various institutions COLLABORATION xxxx --- ----- 1 . . N ---- yyy ----- 1 . . N ---- --- zzzzz Confidentiality : no direct sharing of data Sharing of information granules 1 . . N Vertical collaboration Databases about various clients at several institutions xxxx 1 . . N1 yyyy 1 . . N2 zzzz 1 . . Np DB-1 DB-2 DB-P Horizontal collaboration: Fuzzy C-Means - Identify patterns by optimising local objective functions in each data set ii N k 1 jj c 2 2 u [ ii ] d ik ik [ii ] i 1 N c u k 1 i 1 2 ik [ jj]d ik2 [ jj] Horizontal collaboration: Fuzzy C-Means - Coordinate partition matrices in data sets ii jj U[jj] [ii,jj] U[ii] Communication and collaboration via partition matrices U[ii], U[jj] jj ii [ii,kk] kk U[kk] Horizontal collaboration: Computational details Modified objective function N Q[ii ] k 1 c P N jj1 jj ii k 1 u [ii ]d [ii ] [ii , jj] i 1 2 ik 2 ik Optimization problem c 2 2 { u [ ii ] u [ jj ]} d ik [ii ] ik ik i 1 collaboration Min Q[ii] subject to U[ii] U c N i 1 k 1 U= =. {u ik [ii ] [0,1] | u ik [ii ] 1 for all k and 0 u ik [ii ] N for i}. Horizontal collaboration: Computational details Unconstrained optimisation of V with respect of partition matrix and Lagrange multipliers V[ii] c u i 1 2 ik [ii]d [ii] 2 ik P c α[ii, jj] {u jj1 jj ii ik i 1 The necessary conditions V[ii] 0, u st [ii] V[ii] 0 λ 2 ik c u [ii] u ik [jj]} d [ii] λ( 2 i 1 ik [ii] 1) Horizontal collaboration: Computational details p V[ii] 2 2u st [ii]d st [ii] 2 α[ii, jj](u st [ii] u st [jj])d st2 [ii] λ 0 u st [ii] jj1 jj ii p 2 λ 2d st [ii] α[ii, jj]u st [jj] jjjj1ii u st [ii] p 2 2d st [ii] 1 α[ii, jj]u st [jj] jjjj1ii Horizontal collaboration: Computational details P st [ii] α[ii, jj]u st [jj] jj1 jj ii ψ[ii] P α[ii, jj] jj1 jj ii c u s 1 st [ii] 1 c 1 st [ii] 1 ψ[ii] λ c 1 2 s 1 2d st [ii](1 ψ[ii]) s 1 Horizontal collaboration: Partition matrix st [ii] ust [ii] 1 [ii] c [ii ] 1 jt [1 ] 2 c d st j 1 1 [ii ] d2 j 1 jt Horizontal collaboration: Prototypes N 2 A st [ii] u sk [ii]x kt [ii] A [ii ] C [ii ] v [ii ] B [ii ] D [ii ] st st s s k 1 st N 2 B s [ii] u sk [ii] k 1 C st [ii] P N jj1 jj ii k 1 2 α[ii, jj] (u [ii] u [jj]) x kt [ii ] sk sk D s [ii] P N jj1 jj ii k 1 2 α[ii, jj] (u [ii] u [jj]) sk sk Quantification of the effect of collaboration 1 N c δ u ik [1] u ik [2] | N * c k 1 i 1 Average distance between the partition matrices Quantification of the effect of collaboration 2 -ref 1 -ref 1 1 2 2 1 N c Δ1 u ik [1] u ik [1 ref] | N * c k 1 i 1 1 N c Δ2 u ik [2] u ik [2 ref] | N * c k 1 i 1 Average distance of the partition matrices obtained with and without collaboration