A Logic Programming library for Clojure. At its heart is an original implementation of miniKanren as described in William Byrd’s dissertation Relational Programming in miniKanren: Techniques, Applications, and Implementations. It’s also described in great detail in the The Reasoned Schemer. However, do note that the version that appears in The Reasoned Schemer is an earlier implementation and differs from the one on which this library is based.
Performance is a central concern of this project. Anything that makes it slower will probably not be adopted. Anything that makes it faster without overly complicating the implementation will be considered. It would be interesting to see how we fare on the standard Prolog benchmarks. Currently, on my machine, solving the classic Zebra puzzle 1000 times takes SWI-Prolog about 6 seconds, it takes
logos.minikanren ~2.1s without
A classic AI program:
(ns bratko-logos.monkey-banana (:require [logos.minikanren :as mk] [logos.match :as m])) (m/defne moveo [before action after] ([[:middle :onbox :middle :hasnot] :grasp [:middle :onbox :middle :has]]) ([[?pos :onfloor ?pos ?has] :climb [?pos :onbox ?pos ?has]]) ([[?pos1 :onfloor ?pos1 ?has] :push [?pos2 :onfloor ?pos2 ?has]]) ([[?pos1 :onfloor ?box ?has] :walk [?pos2 :onfloor ?box ?has]])) (m/defne cangeto [state out] ([[_ _ _ :has] true]) ([_ _] (mk/exist [action next] (moveo state action next) (cangeto next out)))) (mk/run 1 [q] (cangeto [:atdoor :onfloor :atwindow :hasnot] q)) ; (true)
A classic Prolog program:
append ( ,Y,Y). append([A|D],Y2,[A|R]) :− append(D,Y2,R).
The Logos version is almost equally succinct:
(defne appendo [x y z] ([() _ y]) ([[?a . ?d] _ [?a . ?r]] (appendo ?d y ?r)))
Logos as of version 0.5.4 supports tabling. Certain kinds of logic programs that would not terminate in Prolog will terminate in Logos if you create a tabled goal.
(defne arco [x y] ([:a :b]) ([:b :a]) ([:b :d])) (def patho (tabled [x y] (conde ((arco x y)) ((exist [z] (arco x z) (patho z y)))))) ;; (:b :a :d) (run* [q] (patho :a q))
Logos supports disequality constraints.
(run* [q] (exist [x y] (!= [x 2] [y 1]) (== x 1) (== y 3) (== q [x y]))) ; ([1 3]) (run* [q] (exist [x y] (!= [x 2] [y 1]) (== x 1) (== y 2) (== q [x y]))) ; ()
Logos comes with a unifier that can be used much like core.unify:
(unifier' '(?x ?y ?z) '(1 2 ?y)) ; (1 2 _.0)
The above is quite slow since we have to walk the data structure and replace the logic var symbols. It’s more efficient to
prep the expressions before hand if you’re going to be unifying the same expressions over and over again.
(let [[u w] (map prep ['(?x ?y) (1 2)])] (unifier u w))
Sometimes it’s useful to create a list of facts that you want to run queries over. Use
fact. Facts are just tuples and Logos will index them by the first element.
(defrel man p) (fact man 'Bob) (fact man 'John) (fact man 'Ricky) (defrel woman p) (fact woman 'Mary) (fact woman 'Martha) (fact woman 'Lucy) (defrel likes p1 p2) (fact likes 'Bob 'Mary) (fact likes 'John 'Martha) (fact likes 'Ricky 'Lucy) (defrel fun p) (fact fun 'Lucy) (run* [q] (exist [x y] (fun y) (likes x y) (== q [x y]))) ; ([Ricky Lucy])
This library is under heavy development as I cover the ideas in Byrd’s thesis and other sources on logic programming. It currently only supports the Clojure 1.3.0 alphas.
This is not the first implementation of miniKanren in Clojure. Jim Duey’s version can be found here. His work on that got me interested in logic programming in the first place.
The core Prolog aspect of Logos is nearing completion. The following are tentative future directions:
- Negation – Stratified Negation as provided by XSB ?
- Constraint Logic Programming – Constraint Handling Rules (CHR) is particularly inspiring
YourKit has has given me a free license for their profiler, greatly simplifying the profiling of Logos performance.
YourKit is kindly supporting open source projects with its full-featured Java Profiler. YourKit, LLC is the creator of innovative and intelligent tools for profiling Java and .NET applications. Take a look at YourKit’s leading software products:
I stayed pretty true to the ideas of the original implementation. There are however several key differences. Unification uses protocols in order leverage the full speed of the host. Clojure’s cons operator differs significantly from Scheme’s so I added the
LConsSeq protocol. Sequences which end in a logic variables can be represented by using
(lcons 'a (lvar 'b)) ; (a . <lvar:b>)
Logic variables are instances of
LVar not vectors as they are in Scheme.
The goal and goal constructor portion has been written from scratch on top of the protocols and makes liberal use of lazy sequences to prevent stack overflow.
Substitutions deftype uses
clojure.lang.PersistentHashMap internally. This may be replaced with something that provides better performance for triangular substitutions.
- Simplicity. Optimizations should not destroy the ideas behind the original design. Any person willing to take take the time to understand the original Scheme implementation should have little trouble understanding how Logos is put together.
- Performance. This implementation is faster than miniKanren under Racket and seems to be close to performnace of miniKanren recorded by the original designers when running under Chez Scheme.
- Emphasis on pure relational programming.
I will only take contributions from people who have submitted Clojure CAs.
- Can we avoid unification? Instead use map/filter when possible which are 10X faster?
- Compilation to pure optimized Clojure ?
- Compilation to byte-code ?
- Fork/join ?
- Efficient Constraint Propagation Engines
- Techniques for Efficient Constraint Propagation
- Operations Research Tools developed at Google
- Solving Every Sudoku Puzzle
- Constraint Handling Rules
- The XSB System Version 3.2 – Volume 2: Libraries, Interfaces, and Packages, particularly the section on Attributed Variables
- The XSB System Version 3.2 – Volume 1: Programmer’s Manual
- Concepts, Technqiues, and Models of Computer Programming, Chapters 9 and 12
- Art of the Propagator
- Constraint Propagation – Models, Techniques, Implementation
- Relational Programming in miniKanren: Techniques, Applications, and Implementations
- The Reasoned Schemer
- Efficient representations for triangular substitutions: A comparison in miniKanren
- A pattern matcher for miniKanren, or, how to get into trouble with CPS macros
- Logical JVM: Implementing the miniKanren logic system in Scala
- Purely Functional Data Strucutres
- Using Datalog with Binary Decision Diagrams for Program Analysis
- Memoing for Logic Programs
- Efficient bottom-up abstract interpretation of prolog by means of constraint solving over symbolic finite domains
Copyright © 2010 David Nolen
Distributed under the Eclipse Public License, the same as Clojure.