## Higher-Order and Modal Logic as a Framework for Explanation-Based Generalization (1989)

### Cached

### Download Links

Citations: | 16 - 6 self |

### BibTeX

@TECHREPORT{Dietzen89higher-orderand,

author = {Scott Dietzen and Frank Pfenning},

title = {Higher-Order and Modal Logic as a Framework for Explanation-Based Generalization},

institution = {},

year = {1989}

}

### Years of Citing Articles

### OpenURL

### Abstract

Logic programming provides a uniform framework in which all aspects of explanation-based generalization and learning may be defined and carried out, but first-order Horn logic is not well suited to application domains such as theorem proving or program synthesis where functions and predicates are the objects of computation. We explore the use of a higher-order representation language and extend EBG to a higher-order logic programming language. Variables may now range over functions and predicates, which leads to an expansion of the space of possible generalizations. We address this problem by extending the logic with the modal ⊔ ⊓ operator (indicating necessary truth) which leads to the language λ ⊔ ⊓ Prolog. We develop a meta-interpreter realizing EBG for λ ⊔ ⊓ Prolog and give some examples in an expanded version of this extended abstract which is available as a technical report [2]. 1

### Citations

757 |
SOAR: An architecture for general intelligence
- Laird, Newell, et al.
- 1987
(Show Context)
Citation Context ...nt, rather than the architecture, is in a position to control assimilation. This approach stands in sharp contrast to systems such as SOAR in which learning is confined to the underlying architecture =-=[28, 41]-=-. Interaction and EBG. Explanation-based generalization is often labeled ‘speed-up’ learning in that EBG extends the domain theory by constructing new rules in the deductive closure of that domain the... |

710 |
An Introduction to Modal Logic
- Hughes, Cresswell
- 1972
(Show Context)
Citation Context ...r EBG. 3 EBG and Modal Logic When describing domain theory and training instance, we distinguish the former with the ⊔⊓ operator. ⊔⊓ is borrowed from modal logic (see, for example, Hughes & Cresswell =-=[6]-=-). It precedes necessarily true sentences, or equivalently, those true in all worlds; non-prefixed sentences are only contingently true, or true in the current world. Previous EBG implementations have... |

696 | A Framework for Defining Logics
- Harper, Honsel, et al.
- 1993
(Show Context)
Citation Context ... (λy. x<y)) . The use of the single name-binding operator λ allows the following operations to be implemented once within the representation domain rather than within the rules of each of its clients =-=[39, 20]-=-: • α-conversion — the renaming of bound variables (e.g., λx.x = λy.y) • β-conversion — capture-avoiding substitution (e.g., (λx.λy.fxy)y = λy ′ .fyy ′ ) • η-conversion — a weak extensionality princip... |

570 | A Transformation System for Developing Recursive Programs
- Burstall, Darlington
- 1976
(Show Context)
Citation Context ... a further discussion of tactic-style theorem proving, see, for example, LCF [17, 37] or Nuprl [4]. 6 Program Development One paradigm for formal program development is that of program transformation =-=[3, 25, 42, 14, 43]-=-. Under such an approach, an abstract specification of an algorithm is refined, or specialized, through a sequence of formal elaboration steps, or transformations, into a program with acceptable perfo... |

490 |
Explanation-Based Generalization: A Unifying View
- Mitchell, Keller, et al.
- 1986
(Show Context)
Citation Context ... is available as a technical report [2]. 1 Introduction To date, work on the explanation-based approach to learning relies primarily on explanation-based generalization (EBG) as its central mechanism =-=[8, 1]-=-. Recently, the logic programming paradigm has been used as a foundation for supporting EBG [7, 11, 5]. One argument put forward in favor of the logic programming framework is that it admits a uniform... |

374 | Uniform Proofs as a Foundation for Logic Programming
- Miller, Nadathur, et al.
- 1991
(Show Context)
Citation Context ...odal Horn clauses to permit higher-order hereditary Harrop formulas [31] (i.e., λProlog) in unboxed contexts. This can be proven by a combination of the methods of del Cerro [6, 7] and Miller, et al. =-=[30]-=-. Without giving a precise definition here, we conjecture a similar result for a more general formulation in which ⊔⊓ may be applied to arbitrary λProlog constructs. The Barcan Formula. The Barcan for... |

306 |
Inductive inference: theory and methods
- Angluin, Smith
- 1983
(Show Context)
Citation Context ...rning and generalization (SBG), on the other hand, rely upon multiple training examples (often both positive and negative) to arrive at an articulation of the sharing among those (positive) instances =-=[1, 8]-=-. While the proof-based generalizations of EBG are necessarily valid, similarity-based generalizations are guaranteed only to the extent that they cover the given examples. Hybrids of these approaches... |

305 |
Quantitative results concerning the utility of explanationbased learning
- Minton
- 1988
(Show Context)
Citation Context ...is assimilation may be selective or may involve the forgetting of those derived rules only infrequently referenced. For a discussion of these issues see Prieditis & Mostow [40, pages 496–497], Minton =-=[33]-=-, and Donat & Wallen [11]. The calculus integration example of Section 5 reinforces our belief that EBG should be a feature of the language rather than a ‘black box’ within the architecture. Consider ... |

302 |
Higher-order abstract syntax
- Pfenning, Elliot
- 1988
(Show Context)
Citation Context ...ny other domains naturally involve binding constructs and are thus best expressed in a higher-order representation domain: logics (when viewed as domains), programming languages, and natural language =-=[10, 9]-=-. Similar representational needs arise when one wants to reason “at the meta-level”, that is, about control strategies for logic programming, theorem proving, or the EBG algorithm itself: One would li... |

152 |
Explanation-based learning: A problem solving perspective
- Minton, Carbonell, et al.
- 1989
(Show Context)
Citation Context ...inductive, or similarity-based learning, and analytical, or explanation-based learning. To date, work on the latter relies primarily on explanation-based generalization (EBG) as its central mechanism =-=[35, 5, 34, 13]-=-. Through the analysis of a formal problem solution (i.e., a proof or explanation), EBG determines the preconditions sufficient to apply the same solution strategy in general. EBG yields a derived rul... |

147 |
Proving and applying program transformations expressed with second-order patterns
- Huet, Lang
- 1978
(Show Context)
Citation Context ... a further discussion of tactic-style theorem proving, see, for example, LCF [17, 37] or Nuprl [4]. 6 Program Development One paradigm for formal program development is that of program transformation =-=[3, 25, 42, 14, 43]-=-. Under such an approach, an abstract specification of an algorithm is refined, or specialized, through a sequence of formal elaboration steps, or transformations, into a program with acceptable perfo... |

114 |
A unification algorithm for typed λ-calculus
- Huet
- 1975
(Show Context)
Citation Context ...s and predicates [36]. Although λProlog is a typed language, we will omit type declarations from our examples for simplicity. λProlog utilizes Huet’s complete algorithm for higher-order 7unification =-=[24]-=-. Although higher-order unification is only semi-decidable and can be highly nondeterministic, Huet’s algorithm is very effective in practice. 4 Modal Logic and EBG EBG relies upon the separation of d... |

99 | An Overview of λProlog
- Nadathur, Miller
- 1988
(Show Context)
Citation Context ...ny other domains naturally involve binding constructs and are thus best expressed in a higher-order representation domain: logics (when viewed as domains), programming languages, and natural language =-=[10, 9]-=-. Similar representational needs arise when one wants to reason “at the meta-level”, that is, about control strategies for logic programming, theorem proving, or the EBG algorithm itself: One would li... |

70 |
Specifying theorem provers in a higherorder logic programming language
- Felty, Miller
- 1988
(Show Context)
Citation Context ...its support for higherorder objects, and in its inclusion of implication and quantification. This promotes its use as a meta-logic for theorem proving. Felty & Miller present this case in more detail =-=[16, 15]-=-. Example 2. The previous integration example relied upon λProlog search to solve queries. Additional levels of control (to constrain search) need not, however, interfere with the underlying EBG proce... |

56 |
Constable et al. Implementing Mathematics with the Nuprl Proof Development System
- Robert
- 1986
(Show Context)
Citation Context ...(2 + 1))) and the generalization boxpi (intgr X\(B * expn X C) X\(B * (expn X (C + 1)) div (C + 1))). For a further discussion of tactic-style theorem proving, see, for example, LCF [17, 37] or Nuprl =-=[4]-=-. 6 Program Development One paradigm for formal program development is that of program transformation [3, 25, 42, 14, 43]. Under such an approach, an abstract specification of an algorithm is refined,... |

55 |
Explanation-based generalization as resolution theorem proving
- Kedar-Cabelli, McCarty
- 1987
(Show Context)
Citation Context ...roach to learning relies primarily on explanation-based generalization (EBG) as its central mechanism [8, 1]. Recently, the logic programming paradigm has been used as a foundation for supporting EBG =-=[7, 11, 5]-=-. One argument put forward in favor of the logic programming framework is that it admits a uniform representation for all aspects of EBG: domain theory, training instance, goal, goal concept, operatio... |

55 |
Explanation-based learning: a survey of programs and perspectives
- Ellman
- 1989
(Show Context)
Citation Context ...inductive, or similarity-based learning, and analytical, or explanation-based learning. To date, work on the latter relies primarily on explanation-based generalization (EBG) as its central mechanism =-=[35, 5, 34, 13]-=-. Through the analysis of a formal problem solution (i.e., a proof or explanation), EBG determines the preconditions sufficient to apply the same solution strategy in general. EBG yields a derived rul... |

49 | A logic programming approach to manipulating formulas and programs
- Miller, Nadathur
- 1987
(Show Context)
Citation Context ...ly involve name binding constructs, and are thus best represented in a higher-order language: logics (when viewed as an object language to be manipulated), programming languages, and natural language =-=[39, 32, 29]-=-. This same lack of adequate representation also arises when one wants to reason ‘at the meta-level’ — that is, about control strategies for logic programming, theorem proving, or the EBG algorithm it... |

46 | Specifying and Implementing Theorem Provers in a Higher-Order Logic Programming Language
- Felty
- 1989
(Show Context)
Citation Context ...its support for higherorder objects, and in its inclusion of implication and quantification. This promotes its use as a meta-logic for theorem proving. Felty & Miller present this case in more detail =-=[16, 15]-=-. Example 2. The previous integration example relied upon λProlog search to solve queries. Additional levels of control (to constrain search) need not, however, interfere with the underlying EBG proce... |

44 | A Survey and Classification of Some Program Transformation Techniques
- Feather
- 1986
(Show Context)
Citation Context ... a further discussion of tactic-style theorem proving, see, for example, LCF [17, 37] or Nuprl [4]. 6 Program Development One paradigm for formal program development is that of program transformation =-=[3, 25, 42, 14, 43]-=-. Under such an approach, an abstract specification of an algorithm is refined, or specialized, through a sequence of formal elaboration steps, or transformations, into a program with acceptable perfo... |

30 |
L.: “Program Improvement by Internal Specialization
- Scherlis
- 1981
(Show Context)
Citation Context |

28 |
Hereditary Harrop formulas and uniform proof systems
- Miller, Nadathur, et al.
- 1987
(Show Context)
Citation Context ...rovided. The formulation of λ ⊔⊓ Prolog we give here is sound and (non-deterministically) complete for an extension of first-order modal Horn clauses to permit higher-order hereditary Harrop formulas =-=[31]-=- (i.e., λProlog) in unboxed contexts. This can be proven by a combination of the methods of del Cerro [6, 7] and Miller, et al. [30]. Without giving a precise definition here, we conjecture a similar ... |

24 |
Automated Proof Search in Non-Classical Logics: Efficient Matrix Methods for Modal and Intuitionistic Logics
- Wallen
- 1990
(Show Context)
Citation Context ...proach incorporating modal logic into the logic programming framework which is independent of EBG [6, 7]. For treatments of automated theorem proving in modal logics outside of logic programming, see =-=[45, 44]-=-. λ⊔⊓ Prolog. As has already been suggested, the motivation for the extended language, λ⊔⊓ Prolog, is that higher-order EBG may be realized within its underlying architecture. λ⊔⊓ Prolog is currently ... |

23 | Some uses of higher-order logic in computational linguistics
- Miller, Nadathur
- 1986
(Show Context)
Citation Context ...ly involve name binding constructs, and are thus best represented in a higher-order language: logics (when viewed as an object language to be manipulated), programming languages, and natural language =-=[39, 32, 29]-=-. This same lack of adequate representation also arises when one wants to reason ‘at the meta-level’ — that is, about control strategies for logic programming, theorem proving, or the EBG algorithm it... |

17 |
Combining empirical and analytical learning with version spaces
- Hirsh
- 1989
(Show Context)
Citation Context ...BG are necessarily valid, similarity-based generalizations are guaranteed only to the extent that they cover the given examples. Hybrids of these approaches are a current topic of research; see Hirsh =-=[21]-=-, for example. Generalization and learning performance are intimately tied to the underlying language for representation, or representation domain. If knowledge is encoded in an inappropriate represen... |

14 |
Explanation-Based Generalization in a Logic-Programming Environment
- Hirsh
- 1987
(Show Context)
Citation Context ...roach to learning relies primarily on explanation-based generalization (EBG) as its central mechanism [8, 1]. Recently, the logic programming paradigm has been used as a foundation for supporting EBG =-=[7, 11, 5]-=-. One argument put forward in favor of the logic programming framework is that it admits a uniform representation for all aspects of EBG: domain theory, training instance, goal, goal concept, operatio... |

13 |
Explanation-Based Generalization: A Alternative View
- DeJong, Mooney
- 1986
(Show Context)
Citation Context ... is available as a technical report [2]. 1 Introduction To date, work on the explanation-based approach to learning relies primarily on explanation-based generalization (EBG) as its central mechanism =-=[8, 1]-=-. Recently, the logic programming paradigm has been used as a foundation for supporting EBG [7, 11, 5]. One argument put forward in favor of the logic programming framework is that it admits a uniform... |

9 |
Tactics and Tacticals in Cambridge LCF
- Paulson
- 1979
(Show Context)
Citation Context ...xpn X (2 + 1) div (2 + 1))) and the generalization boxpi (intgr X\(B * expn X C) X\(B * (expn X (C + 1)) div (C + 1))). For a further discussion of tactic-style theorem proving, see, for example, LCF =-=[17, 37]-=- or Nuprl [4]. 6 Program Development One paradigm for formal program development is that of program transformation [3, 25, 42, 14, 43]. Under such an approach, an abstract specification of an algorith... |

9 |
Automated Theorem-Proving in Non-Classical Logics
- thistlewaite, mcrobbie, et al.
- 1988
(Show Context)
Citation Context ...proach incorporating modal logic into the logic programming framework which is independent of EBG [6, 7]. For treatments of automated theorem proving in modal logics outside of logic programming, see =-=[45, 44]-=-. λ⊔⊓ Prolog. As has already been suggested, the motivation for the extended language, λ⊔⊓ Prolog, is that higher-order EBG may be realized within its underlying architecture. λ⊔⊓ Prolog is currently ... |

8 |
Reasoning about operationality for explanation-based learning
- Hirsh
- 1988
(Show Context)
Citation Context ...log supports both dynamic and goal-based operationality criteria within the same uniform framework of higher-order logic. (This is similarly possible in Prolog through its meta-programming facilities =-=[22, 23]-=-.) Incorporating operationality into our implementation requires providing the meta-interpreter with access to an operationality predicate. The revision, which is illustrated in Figure 13, involves th... |

6 |
Learning and applying generalised solutions using higher order resolution
- Donat, Wallen
- 1988
(Show Context)
Citation Context ...guage supports EBG, and largely ignore questions about how these generalizations may be assimilated and applied automatically. For a discussion of these issues, see Prieditis [11], and Donat & Wallen =-=[4]-=-. We also do not deal with the important concept of operationality; see Hirsh [5] for a treatment in the context of logic programming. We begin by motivating the step to a higher-order language as the... |

4 |
Analogy in program development
- Dietzen, Scherlis
- 1987
(Show Context)
Citation Context ...hodology, we will generalize over a simple transformational system which we have applied to induce tail recursion in certain situations. This example is also treated tentatively in Dietzen & Scherlis =-=[10]-=-, among others. From a tail recursive version, an iterative form could easily be derived. It is not our intention that the example be grasped in detail; rather we present it for the interesting genera... |

4 |
Fariñas del Cerro. MOLOG: A system that extends PROLOG with modal logic
- Luis
- 1986
(Show Context)
Citation Context ...the inclusion of an implication within the preconditions of a derived rule. Del Cerro takes another approach incorporating modal logic into the logic programming framework which is independent of EBG =-=[6, 7]-=-. For treatments of automated theorem proving in modal logics outside of logic programming, see [45, 44]. λ⊔⊓ Prolog. As has already been suggested, the motivation for the extended language, λ⊔⊓ Prolo... |

4 |
eLP: A Common Lisp implementation of λProlog in the Ergo Support System. Available via ftp over the Internet
- Elliott, Pfenning
- 1989
(Show Context)
Citation Context ...cknowledgments. Our meta-interpreter depends on the eLP, the implementation of λProlog developed by Conal Elliott and Frank Pfenning in the framework of the Ergo project at Carnegie Mellon University =-=[12]-=-. We thank Conal Elliott, Masami Hagiya, Haym Hirsh, Dale Miller, Tom Mitchell, and William Scherlis for their thoughtful comments on our presentation. 2 First-order EBG in the Logic Programming Frame... |

3 | Generalization from partial parameterization in higher-order type theory. Theoretical Computer Science, 63:113–139 - Hagiya - 1989 |

3 |
Knowledge-based software development tools
- Smith, Pressburger
- 1988
(Show Context)
Citation Context |

2 |
Prieditis and Jack Mostow. Prolearn: Toward a Prolog interpreter that learns
- Armand
- 1987
(Show Context)
Citation Context ...roach to learning relies primarily on explanation-based generalization (EBG) as its central mechanism [8, 1]. Recently, the logic programming paradigm has been used as a foundation for supporting EBG =-=[7, 11, 5]-=-. One argument put forward in favor of the logic programming framework is that it admits a uniform representation for all aspects of EBG: domain theory, training instance, goal, goal concept, operatio... |

2 |
A correctness proof of explanation-based generalization as resolution theorem proving
- Bhatnagar
- 1988
(Show Context)
Citation Context ...is, EBG in which functions and predicates as well as firstorder constants may be abstracted, or replaced with variables. Recently, the logic programming paradigm has been used as a foundation for EBG =-=[27, 40, 22, 2]-=-. One argument put forward in favor of the logic programming framework is that it admits a uniform representation for all aspects of EBG: domain theory, training instance, goal, derived rule, operatio... |

2 |
Fariñas del Cerro and Martti Penttonen. A note on the complexity of the satisfiability of modal Horn Clauses
- Luis
- 1987
(Show Context)
Citation Context ...the inclusion of an implication within the preconditions of a derived rule. Del Cerro takes another approach incorporating modal logic into the logic programming framework which is independent of EBG =-=[6, 7]-=-. For treatments of automated theorem proving in modal logics outside of logic programming, see [45, 44]. λ⊔⊓ Prolog. As has already been suggested, the motivation for the extended language, λ⊔⊓ Prolo... |

2 |
Mapping. Explanation-Based Generalization into Soar
- Rosenbloom, Laird
- 1986
(Show Context)
Citation Context ...nt, rather than the architecture, is in a position to control assimilation. This approach stands in sharp contrast to systems such as SOAR in which learning is confined to the underlying architecture =-=[28, 41]-=-. Interaction and EBG. Explanation-based generalization is often labeled ‘speed-up’ learning in that EBG extends the domain theory by constructing new rules in the deductive closure of that domain the... |

1 |
Dietterich et al. Learning and inductive inference
- M
- 1982
(Show Context)
Citation Context ...rning and generalization (SBG), on the other hand, rely upon multiple training examples (often both positive and negative) to arrive at an articulation of the sharing among those (positive) instances =-=[1, 8]-=-. While the proof-based generalizations of EBG are necessarily valid, similarity-based generalizations are guaranteed only to the extent that they cover the given examples. Hybrids of these approaches... |