## A Strategy for Constructing New Predicates in First Order Logic (1988)

Venue: | In Proceedings of the Third European Working Session on Learning |

Citations: | 16 - 6 self |

### BibTeX

@INPROCEEDINGS{Muggleton88astrategy,

author = {Stephen Muggleton},

title = {A Strategy for Constructing New Predicates in First Order Logic},

booktitle = {In Proceedings of the Third European Working Session on Learning},

year = {1988},

pages = {123--130},

publisher = {Pitman}

}

### OpenURL

### Abstract

There is increasing interest within the Machine Learning community in systems which automatically reformulate their problem representation by defining and constructing new predicates. A previous paper discussed such a system, called CIGOL, and gave a derivation for the mechanism of inverting individual steps in first order resolution proofs. In this paper we describe an enhancement to CIGOL's learning strategy which strongly constrains the formation of new concepts and hypotheses. The new strategy is based on results from algorithmic information theory. Using these results it is possible to compute the probability that the simplifications produced by adopting new concepts or hypotheses are not based on chance regularities within the examples. This can be derived from the amount of information compression produced by replacing the examples with the hypothesised concepts. CIGOL's improved performance, based on an approximation of this strategy, is demonstrated by way of the automatic "di...

### Citations

6041 |
A mathematical theory of communication
- Shannon
- 1948
(Show Context)
Citation Context ...exity. The basic intuition rests on the observation that although the strings 010100110111001100010110101100 and 010101010101010101010101010101 have approximately the same Shannon information content =-=[25]-=-, the second contains a higher degree of regularity than the first. As an alternative to standard information measures Kolmogorov defined the algorithmic information of a finite string s as being equa... |

945 |
A machine-oriented logic based on the resolution principle
- Robinson
- 1965
(Show Context)
Citation Context ...uctive rules of inference. However, we note that this is somewhat analogous to the position in theorem proving before the introduction of the universal rule of deductive inference known as resolution =-=[23]-=-. The thesis behind CIGOL is that appropriate inversions of resolution provide an efficient, sufficient and complete mechanism for the inductive setting described by (1). Question 2 Given that the pre... |

670 |
A theory and methodology of inductive learning
- Michalski
- 1983
(Show Context)
Citation Context ...ruct H given I, B, and E + ? Answer: All methods outside enumeration and testing of H rely on applying efficient "generalisation operations" which incrementally construct H from B and E + . =-=Michalski [12]-=- notes that these generalisation operations can be based on reversing the deductive rules of inference which allow us to derive E + from B and H. Michalski's INDUCE system uses the inversions of a wid... |

552 |
Generalization as Search
- Mitchell
- 1982
(Show Context)
Citation Context ...dge B and irrelevant background knowledge I? In the following sections we discuss possible approaches to answering questions 3-5. 3 Search strategies and algorithmic information 3.1 Version spaces In =-=[14] Mitchell -=-describes a general search strategy for inductive inference, known as the "Version space" approach. This method involves the maintenance of two sets, S and G. These sets represent respective... |

521 |
Three approaches to the quantitive definition of information
- Kolmogorov
- 1965
(Show Context)
Citation Context ...ch a lattice. We must therefore look to some alternative model to guide and constrain the search through this more complex lattice. 3.2 Algorithmic information theory Following the lead of Kolmogorov =-=[8]-=- various information theorists [4, 26, 2] have investigated the relationship between computation, randomness and message complexity. The basic intuition rests on the observation that although the stri... |

490 |
Explanation-Based Generalization: A Unifying View
- Mitchell, Keller, et al.
- 1986
(Show Context)
Citation Context ... out 1. inductive concept formation such as [11, 21] 2. constructive induction such as [22, 13] 3. discovery such as [10, 9, 6] 4. generalisation of single examples using background knowledgessuch as =-=[5, 15, 24]-=- Unlike most learning systems described in the literature CIGOL uses an unrestricted form of first order Horn clause logic which allows predicate relations to take not only variables and constants as ... |

404 |
A formal theory of inductive inference
- Solomonoff
- 1964
(Show Context)
Citation Context ...ok to some alternative model to guide and constrain the search through this more complex lattice. 3.2 Algorithmic information theory Following the lead of Kolmogorov [8] various information theorists =-=[4, 26, 2]-=- have investigated the relationship between computation, randomness and message complexity. The basic intuition rests on the observation that although the strings 010100110111001100010110101100 and 01... |

280 |
Conjectures and Refutations: The Growth of Scientific Knowledge. Third edition. London: Routledge and Kegan
- Popper
- 1969
(Show Context)
Citation Context ...nishing returns in the confidence inspired in a theory involved in showing that it covers an increasingly large number of facts. All of these are standard assumptions within the philosophy of science =-=[20]-=-. It is now necessary to describe in more detail the reference Turing machinesT r and its input and output tape encodings I and O. 3.4 Encodings and the compression model First an efficient Turing tap... |

246 |
Learning from observation: conceptual clustering
- Michalski, Stepp
- 1983
(Show Context)
Citation Context ...llows CIGOL to exhibit a number of facets of Machine Learning. Thus CIGOL can be classed with systems which carry out 1. inductive concept formation such as [11, 21] 2. constructive induction such as =-=[22, 13]-=- 3. discovery such as [10, 9, 6] 4. generalisation of single examples using background knowledgessuch as [5, 15, 24] Unlike most learning systems described in the literature CIGOL uses an unrestricted... |

179 |
Machine invention of first-order predicates by inverting resolution
- Muggleton, Buntine
- 1988
(Show Context)
Citation Context ...047. 1 Introduction A concept can only be learned if it can be represented. More than this, an appropriate representation language facilitates a simple and elegant description of a target concept. In =-=[17]-=- we describe a system called CIGOL which automatically develops its own representation language in order to efficiently represent target concepts. Initially CIGOL is provided with pertinent background... |

118 |
On formally undecidable propositions of ‘Principia Mathematica’ and related systems
- Gödel
- 1931
(Show Context)
Citation Context ...ances entailed by P . Now append the numbers corresponding to the particular examples in E + onto the encoded description of P to make I(P ). Such a numbering scheme is called a Godel numbering after =-=[7]-=-. A natural 1 Prefix codes are variable length bit patterns used to encode the symbols in a message. Efficient coding schemes allow one to encode each symbol in close to the optimal of \Gammalog 2 p b... |

98 | Learning concepts by asking questions
- Sammut, Banerji
- 1986
(Show Context)
Citation Context ... out 1. inductive concept formation such as [11, 21] 2. constructive induction such as [22, 13] 3. discovery such as [10, 9, 6] 4. generalisation of single examples using background knowledgessuch as =-=[5, 15, 24]-=- Unlike most learning systems described in the literature CIGOL uses an unrestricted form of first order Horn clause logic which allows predicate relations to take not only variables and constants as ... |

93 |
Generalized Subsumption and its Applications to Induction and Redundancy
- Buntine
- 1988
(Show Context)
Citation Context ...t this simple definition allows us not only to compare the relative generality of atomic formulae and clauses but also the same relationship for arbitrary pairs of theories (sets of clauses). Buntine =-=[3] describes-=- an algorithm aimed at computing this generality relationship which he terms "generalised subsumption". For a fuller discussion of the subject of generality the reader is referred to Niblett... |

79 |
Automatic Methods of Inductive Inference
- Plotkin
- 1971
(Show Context)
Citation Context ...imed at computing this generality relationship which he terms "generalised subsumption". For a fuller discussion of the subject of generality the reader is referred to Niblett [18]. Followin=-=g Plotkin [19]-=- we may more precisely define the setting of inductive learning using the following relationship IsBsH ` E + (1) where I is background knowledge which is not pertinent to the present learning problem,... |

70 |
Selection of the Most Representative Training Examples and Incremental Generation of VL1 Hypotheses: the Underlying Methodology and the
- RS, Larson
- 1978
(Show Context)
Citation Context ...epts. The generality of the approach used allows CIGOL to exhibit a number of facets of Machine Learning. Thus CIGOL can be classed with systems which carry out 1. inductive concept formation such as =-=[11, 21]-=- 2. constructive induction such as [22, 13] 3. discovery such as [10, 9, 6] 4. generalisation of single examples using background knowledgessuch as [5, 15, 24] Unlike most learning systems described i... |

58 |
Rediscovering chemistry with the bacon system
- Langley, Bradshaw, et al.
- 1984
(Show Context)
Citation Context ...r of facets of Machine Learning. Thus CIGOL can be classed with systems which carry out 1. inductive concept formation such as [11, 21] 2. constructive induction such as [22, 13] 3. discovery such as =-=[10, 9, 6]-=- 4. generalisation of single examples using background knowledgessuch as [5, 15, 24] Unlike most learning systems described in the literature CIGOL uses an unrestricted form of first order Horn clause... |

48 | Logical Depth and Physical Complexity
- Bennett
- 1988
(Show Context)
Citation Context ...ok to some alternative model to guide and constrain the search through this more complex lattice. 3.2 Algorithmic information theory Following the lead of Kolmogorov [8] various information theorists =-=[4, 26, 2]-=- have investigated the relationship between computation, randomness and message complexity. The basic intuition rests on the observation that although the strings 010100110111001100010110101100 and 01... |

40 |
Discovering rules from large collections of examples - a case study
- Quinlan
- 1979
(Show Context)
Citation Context ...epts. The generality of the approach used allows CIGOL to exhibit a number of facets of Machine Learning. Thus CIGOL can be classed with systems which carry out 1. inductive concept formation such as =-=[11, 21]-=- 2. constructive induction such as [22, 13] 3. discovery such as [10, 9, 6] 4. generalisation of single examples using background knowledgessuch as [5, 15, 24] Unlike most learning systems described i... |

32 |
Substantial constructive induction using layered information compression: Tractable feature formation in search
- Rendell
- 1985
(Show Context)
Citation Context ...llows CIGOL to exhibit a number of facets of Machine Learning. Thus CIGOL can be classed with systems which carry out 1. inductive concept formation such as [11, 21] 2. constructive induction such as =-=[22, 13]-=- 3. discovery such as [10, 9, 6] 4. generalisation of single examples using background knowledgessuch as [5, 15, 24] Unlike most learning systems described in the literature CIGOL uses an unrestricted... |

29 |
On automated scientific theory formation: A case study using the AM program
- Lenat
- 1977
(Show Context)
Citation Context ...r of facets of Machine Learning. Thus CIGOL can be classed with systems which carry out 1. inductive concept formation such as [11, 21] 2. constructive induction such as [22, 13] 3. discovery such as =-=[10, 9, 6]-=- 4. generalisation of single examples using background knowledgessuch as [5, 15, 24] Unlike most learning systems described in the literature CIGOL uses an unrestricted form of first order Horn clause... |

28 |
An Oracle Based Approach to Constructive Induction
- Duce
- 1987
(Show Context)
Citation Context ...esis forming mechanisms employed by CIGOL are based on inverting individual steps of a resolution proof. This approach is a generalisation of the approaches used by Sammut and Banerji [24], Muggleton =-=[16]-=- and Banerji [1]. Other strongly related work in progress can be found in Wirth [27] and Wrobel [28]. In [17] we provided a derivation for the inverse resolution operators employed by CIGOL but to a l... |

25 |
Generalizations based on explanations
- DeJong
(Show Context)
Citation Context ... out 1. inductive concept formation such as [11, 21] 2. constructive induction such as [22, 13] 3. discovery such as [10, 9, 6] 4. generalisation of single examples using background knowledgessuch as =-=[5, 15, 24]-=- Unlike most learning systems described in the literature CIGOL uses an unrestricted form of first order Horn clause logic which allows predicate relations to take not only variables and constants as ... |

18 |
On the discovery of mathematical theorems
- Epstein
- 1987
(Show Context)
Citation Context ...r of facets of Machine Learning. Thus CIGOL can be classed with systems which carry out 1. inductive concept formation such as [11, 21] 2. constructive induction such as [22, 13] 3. discovery such as =-=[10, 9, 6]-=- 4. generalisation of single examples using background knowledgessuch as [5, 15, 24] Unlike most learning systems described in the literature CIGOL uses an unrestricted form of first order Horn clause... |

17 | A study of generalisation . in logic programs - Niblett - 1988 |

17 |
Automatic representation adjustment in an observational discovery system
- Wrobel
- 1988
(Show Context)
Citation Context ...oof. This approach is a generalisation of the approaches used by Sammut and Banerji [24], Muggleton [16] and Banerji [1]. Other strongly related work in progress can be found in Wirth [27] and Wrobel =-=[28]-=-. In [17] we provided a derivation for the inverse resolution operators employed by CIGOL but to a large degree left open the question of strategy of operator application. The result was that CIGOL as... |

11 |
Learning by failure to prove
- Wirth
- 1988
(Show Context)
Citation Context ... a resolution proof. This approach is a generalisation of the approaches used by Sammut and Banerji [24], Muggleton [16] and Banerji [1]. Other strongly related work in progress can be found in Wirth =-=[27]-=- and Wrobel [28]. In [17] we provided a derivation for the inverse resolution operators employed by CIGOL but to a large degree left open the question of strategy of operator application. The result w... |

5 |
Learning in the limit in a growing language
- Banerji
- 1987
(Show Context)
Citation Context ...anisms employed by CIGOL are based on inverting individual steps of a resolution proof. This approach is a generalisation of the approaches used by Sammut and Banerji [24], Muggleton [16] and Banerji =-=[1]-=-. Other strongly related work in progress can be found in Wirth [27] and Wrobel [28]. In [17] we provided a derivation for the inverse resolution operators employed by CIGOL but to a large degree left... |