Posts Tagged: Abstraction


13
Jan 12

Why do most programmers work so hard at pretending that they’re not doing math?

In the early days programming was considered a subdiscipline of mathematics. In fact, the very first person to write an algorithm was renowned as a mathematical genius. However, somewhere along the way we forgot. We began to think of ourselves as something different, a profession not beholden to rigor or deep understanding of the models we create.

It’s easy to see how this would happen within an industry in which so much more weight is put on knowing the API of the year over understanding base principles or expressing the desire to dig deep. People can make huge amounts of money pushing methodologies when they have little to no evidence of effectiveness. We work in an environment where hearsay and taste drive change instead of studies and models. We are stumbling in the dark.

I have come to attribute our sorry state to a combination of factors. The first is the lack of formal training for most programmers. This isn’t in itself a bad thing, but when combined with a lack of exposure to mathematics beyond arithmetic, due primarily to an inadequate school system, we are left with a huge number of people who think programming and math are unrelated. They see every day how their world is filled with repeating patterns but they miss the beautiful truth that math is really about modeling patterns and that numbers are just a small part of that.

The relationship between math and programming extends far beyond SQL’s foundation in set theory or bits being governed by information theory. Math is intertwined within the code you write every day. The relationships between the different constructs you create and the patterns you use to create them are math too. This is why typed functional programming is so important: it’s only when you formalize these relationships into a model that their inconsistencies become apparent.

Most recently it seems to have become a trend to think of testing as a reasonable substitute for models. Imagine if physics was done this way. What if the only way we knew how to predict how an event would turn out was to actually measure it each time? We wouldn’t be able to generalize our findings to even a slightly different set of circumstances. But it gets even worse: How would you even know what your measurement is of without a model to measure it within? To move back into the context of programming: how useful is a test that doesn’t capture all of the necessary input state?

This is exactly why dynamic programs with mutation become such a mess. Object oriented structure only makes things worse by hiding information and state. Implicit relationships, complex hidden structure and mutation each impair reasoning alone, but when combined they create an explosion of complexity. They each conspire with each other to create a system which defies comprehensive modeling by even the brightest minds.

This is also why programmers who use typed functional languages frequently brag about their low bug count yet eschew methodologies like test driven development: tests pale in comparison to the power of actual models.

Many thanks to my friends on twitter for the discussion leading up to this post: @seanschade, @craigstuntz, @sswistun, @joakinen, @copumpkin, @dibblego, @taylodl, @TheColonial, @tomasekeli, @tim_g_robinson and @CarstKoenig. Special thanks to @danfinch for taking the time to proof read.

19
Sep 11

Imperative Pseudocode to Pure Functional Algorithm with Gale-Shapely and F#

Continuing from last time, let’s look at how one goes from imperative pseudocode to pure functional using Gale-Shapely as an example.

Overall, to convert an algorithm from imperative to functional is a fairly simple process once you understand how to convert from a while loop to recursion with accumulators. This post is just a more advanced version of what I talked about in my From Imperative to Computation Expressions post. In fact, if you haven’t already I suggest you read it first. I’ll be assuming you understand how to perform that conversion in this article.  If you’re still having trouble, another good article which is a bit easier is The Ted Neward F# Folding Challenge.

The conversion of an algorithm is a two stage process: first identify algorithmic cases, implied mutable data structures, and alternate immutable data structures.  Then convert the given loops into recursion while leveraging pattern matching.  I know it may seem difficult at first, but after reading this I think you’ll have a good idea of how to get there.

So, let’s identify the cases.

function stableMatching {   
    Initialize all m ? M and w ? W to free
    // Loop Termination Case on
    // check for “free m with unproposed w” and
 
    // implicit ranking of W for each m in M
   
    while ? free man m who still has a
            woman w to propose to {      
       w = m’s highest ranked such woman
               who he has not proposed to yet
      
// She’s Single Case 
       if w is free
         (m, w) become engaged
       // She’s Engaged Case 
       else some pair (m’, w) already exists
         // She Picks m Sub-Case 

         if w prefers m to m’
           (m, w) become engaged
           m’ becomes free
         // She Picks m’ Sub-Case 
         else
           (m’, w) remain engaged
    }
}

Starting from the top, we can just ignore the imperative initialization.  However, you’ll notice that a great deal is going on in that while clause. 

First, it defines our loop termination by looking for a “free m who still has a w to propose to”.  This is interesting because it seems to be begging for some kind of data structure lookup.  Second it defines a implicit ranking of each member of W by each member of M.  Note that this is not mentioned in the initialization phase.

The rest of this algorithm reads like a simple list of cases.  To make this plain, let’s list them out along with state changes.

  1. no single m in M with unproposed w 
    — execution terminates
  2. m proposes to unengaged w
    — w is no longer in m proposals
    — m is no longer single
    — m and w are engaged
  3. m proposes to w engaged to m’
    a.  She switches to m
          — w is no longer in m proposals
          — m is no longer single
          — m and w are engaged
          — m’ is single
    b.  She stays with m’
          — w is no longer in m proposals
          — m’ and w stay engaged
          — m remains single

The mechanics of what exactly we do in these cases depends on our data structures.  So we can’t get much further without figuring that out.  To identify what data structures are needed we’ll need to identify what exactly it is that needs to be kept track of.  Only once we know that can we pick our immutable data structures appropriately. 

Looking above, it appears that we’ll need to keep track of:

  1. Given a woman, is she engaged and, if so, who is she engaged to? (need to look it up)
  2. Which men are single? (just need one at a time, order doesn’t matter)
  3. For each man, which is his next most desirable match that he has not yet proposed to? (just need one at a time, order matters)

Note that we don’t have to keep track of which men are engaged to which women, the question is never asked.  This is a minimal set of requirements here.  No extra stuff allowed.  I’m serious.

Moving on, in an imperative mutable version you might use these data structures:

  1. An array indexed by w to the m index engaged to (-1 if single)
  2. An array indexed by m to the w index engaged to (-1 if single)
  3. An array indexed by m where each element is an ordered linked list of unproposed w indices.

However, what would satisfy these same requirements but be immutable?  Well, if we step out of the mutation mindset it becomes obvious pretty quickly:

  1. An engagement map of women to men
  2. An immutable linked list of single men
  3. Define a man to include his index as well as an ordered unproposed woman immutable linked list

Instead of scanning arrays, or querying mutable structures we simply decompose our data structures as much as needed and recompose them into the form we need on the next call.   This may sound difficult, but it’s really not.  The fact that we have pattern matching and our cases can be defined in terms of the shape of our data structures actually makes this very simple to do.  That is, once you get the hang of it.

To see this in action, let’s break up the code in the last article and discuss each chunk.

 1: open System
 2: 
 3: // a Bachelor is an identity index and an 
 4: // ordered list of women indicies to approach.
 5: type Bachelor = int * int list

Here we see the definition of a bachelor, it’s a tuple of an integer and a list of integers. This is effectively a pair containing an index and a list of woman indices. Note that in this example all men are of type Bachelor.

 7: // Some notation:
 8: // wi = woman index (int)
 9: // mi = man index (int)
10: // mi' = woman's current partner index (int)
11: // m = man with index and unapproached women indices (Bachelor)
12: // mSingle = men that are single (Bachelor list)
13: // wEngaged = engagements from women to men (int, Bachelor)

We’ll keep this notation around just so you have easy reference.

15: let funGS (comp: _ -> _ -> float) (M: _ array) (W: _ array) =
16:   let Windices = [ 0 .. W.Length - 1 ]
17:   // List of men with women in order of desire  
18:   let Munproposed = 
19:     List.init M.Length 
20:       (fun mi -> 
21:            let sortFun wi = 1.0 - (comp M.[mi] W.[wi])
22:            mi, Windices |> List.sortBy sortFun)

Windices is just a list of numbers from 0 to the number of woman tokens. Doing this first makes the calculation of Munproposed less expensive.

Munproposed is our initial set of all single men. Here List.init is used to call the given lambda to generate each element. As you can see by the last line within that lambda, each element of the list is a tuple containing the index of that man and a list of women indices sorted by the given desirability function, sortFun.

23:   // Recursively solve stable marriages
24:   let rec findMarriages mSingle wEngaged =
25:     match mSingle with
26:     // No single guys left with desired women, we're done
27:     | [] -> wEngaged
28:     // Guy is out of luck, remove from singles
29:     | (mi, []) :: bachelors -> findMarriages bachelors wEngaged

These are our first two cases, if the mSingle list pattern matches with [], we know it’s empty and so we are done.

In the second case we are pattern matching on the structure of singles list as well as it’s first element.  The syntax (mi, []) matches on a tuple of an index and an empty inner list.  This list is the list of yet unproposed to women and so if it’s empty we know that this guy is out of options and so it’s ok to drop him from the list of bachelors.  We do this by simply recursing on bachelors, which is the tail of the mSingle list.

30:     // He's got options!
31:     | (mi, wi :: rest) :: bachelors -> 
32:       let m = mi, rest

This is our general purpose match, and is the structure we expect to see in most cases.  Here the head is fully decomposed into a man index (mi), his next first choice woman index (wi), the rest of the women (rest) and the rest of the single men (bachelors).

Immediately afterward we define m to be that given man index (mi) and the rest of the women tokens (rest).  As two tokens are only ever compared once, m should no longer contain wi in his ordered list of proposals.  This is the form he will take until the algorithm ends or he is evaluated again after being placed in the mSingle pool.

33:       match wEngaged |> Map.tryFind wi with
34:       // She's single! m is now engaged!
35:       | None -> findMarriages bachelors (wEngaged |> Map.add wi m)
36:       // She's already engaged, let the best man win!
37:       | Some (m') –> 
38: let mi', _ = m'

Now that we have fully matched out everything masculine, it’s time to turn to the feminine. 

Here we look in the wEngaged map to find out if wi is single.  This is done by a using Map.tryFind which returns None if the given key is not in the map and Some(value) if it is.

If the Map.tryFind call returns None we know she is single and so recurse with the rest of our single men (bachelors) and our wEngaged map is extended via Map.add to contain a new mapping from the woman index (wi) to the given man (m).

If she’s engaged our pattern match automatically decomposes the given option type and we know exactly who she is engaged to (m’).

39:         if comp W.[wi] M.[mi] > comp W.[wi] M.[mi'] then 
40:           // Congrats mi, he is now engaged to wi
41:           // The previous suitor (mi') is bested 
42:           findMarriages 
43:             (m' :: bachelors) 
44:             (wEngaged |> Map.add wi m)
45:         else
46:           // The current bachelor (mi) lost, better luck next time
47:           findMarriages 
48:             (m :: bachelors) 
49:             wEngaged

This is the core comparison logic.  Here we determine which bachelor gets her hand.  This is pretty straightforward from an imperative perspective as we’re using the oh-so familiar if statement.

In the first case the new contender (m) wins.  We append the loser (m’) back on to the head of what will be mSingle next time around with the syntax (m’ :: bachelors).  Similar to what happens if she’s single, we add a mapping from the woman index (wi) to the man instance (m) via the call to Map.add.  Note that this will override the former mapping from wi to m’.

If the current bachelor loses we simply append him back to the head of the bachelors list and try again next time around.  You’ll get her next time bro.

50:   findMarriages Munproposed Map.empty
51:   // Before returning, remove unproposed lists from man instances  
52:   |> Map.map (fun wi m -> let mi, _ = m in mi)  

The final section is just how we set up our recursion.  We start with the full list of single men and their ranked lady lists (Munproposed) and an empty map of relationships (Map.empty).  When the algorithm exits we make one last pass through our returned wEngaged map in order to strip out all of the lists of unproposed to women.  It’s just baggage now anyway.

Well that’s how you get from imperative pseudocode to a pure functional implementation.  All of the code here, as well as a really nasty imperative version to compare it to is available on my GitHub page

If you liked this post, have any questions, or feel like I missed something important I hope you’ll comment here on it.  We’ll all benefit from the knowledge shared. 

namespace System
type Bachelor = int * int list

Full name: Snippet.Bachelor

Multiple items

val int : ‘T -> int (requires member op_Explicit)

Full name: Microsoft.FSharp.Core.Operators.int

——————–

type int<‘Measure> = int

Full name: Microsoft.FSharp.Core.int<_>

  type: int<‘Measure>

  implements: IComparable

  implements: IConvertible

  implements: IFormattable

  implements: IComparable<int<‘Measure>>

  implements: IEquatable<int<‘Measure>>

  inherits: ValueType

——————–

type int = int32

Full name: Microsoft.FSharp.Core.int

  type: int

  implements: IComparable

  implements: IFormattable

  implements: IConvertible

  implements: IComparable<int>

  implements: IEquatable<int>

  inherits: ValueType

type ‘T list = List<‘T>

Full name: Microsoft.FSharp.Collections.list<_>

  type: ‘T list

  implements: Collections.IStructuralEquatable

  implements: IComparable<List<‘T>>

  implements: IComparable

  implements: Collections.IStructuralComparable

  implements: Collections.Generic.IEnumerable<‘T>

  implements: Collections.IEnumerable

val funGS : (‘a -> ‘a -> float) -> ‘a array -> ‘a array -> Map<int,int>

Full name: Snippet.funGS

val comp : (‘a -> ‘a -> float)
Multiple items

val float : ‘T -> float (requires member op_Explicit)

Full name: Microsoft.FSharp.Core.Operators.float

——————–

type float<‘Measure> = float

Full name: Microsoft.FSharp.Core.float<_>

  type: float<‘Measure>

  implements: IComparable

  implements: IConvertible

  implements: IFormattable

  implements: IComparable<float<‘Measure>>

  implements: IEquatable<float<‘Measure>>

  inherits: ValueType

——————–

type float = Double

Full name: Microsoft.FSharp.Core.float

  type: float

  implements: IComparable

  implements: IFormattable

  implements: IConvertible

  implements: IComparable<float>

  implements: IEquatable<float>

  inherits: ValueType

Multiple items

val M : ‘a array

  type: ‘a array

  implements: ICloneable

  implements: Collections.IList

  implements: Collections.ICollection

  implements: Collections.IStructuralComparable

  implements: Collections.IStructuralEquatable

  implements: Collections.Generic.IList<‘a>

  implements: Collections.Generic.ICollection<‘a>

  implements: seq<‘a>

  implements: Collections.IEnumerable

  inherits: Array

——————–

val M : ‘a array

  type: ‘a array

  implements: ICloneable

  implements: Collections.IList

  implements: Collections.ICollection

  implements: Collections.IStructuralComparable

  implements: Collections.IStructuralEquatable

  implements: Collections.Generic.IList<‘a>

  implements: Collections.Generic.ICollection<‘a>

  implements: seq<‘a>

  implements: Collections.IEnumerable

  inherits: Array

type ‘T array = ‘T []

Full name: Microsoft.FSharp.Core.array<_>

  type: ‘T array

  implements: ICloneable

  implements: Collections.IList

  implements: Collections.ICollection

  implements: Collections.IStructuralComparable

  implements: Collections.IStructuralEquatable

  implements: Collections.Generic.IList<‘T>

  implements: Collections.Generic.ICollection<‘T>

  implements: seq<‘T>

  implements: Collections.IEnumerable

  inherits: Array

Multiple items

val W : ‘a array

  type: ‘a array

  implements: ICloneable

  implements: Collections.IList

  implements: Collections.ICollection

  implements: Collections.IStructuralComparable

  implements: Collections.IStructuralEquatable

  implements: Collections.Generic.IList<‘a>

  implements: Collections.Generic.ICollection<‘a>

  implements: seq<‘a>

  implements: Collections.IEnumerable

  inherits: Array

——————–

val W : ‘a array

  type: ‘a array

  implements: ICloneable

  implements: Collections.IList

  implements: Collections.ICollection

  implements: Collections.IStructuralComparable

  implements: Collections.IStructuralEquatable

  implements: Collections.Generic.IList<‘a>

  implements: Collections.Generic.ICollection<‘a>

  implements: seq<‘a>

  implements: Collections.IEnumerable

  inherits: Array

val Windices : int list

  type: int list

  implements: Collections.IStructuralEquatable

  implements: IComparable<List<int>>

  implements: IComparable

  implements: Collections.IStructuralComparable

  implements: Collections.Generic.IEnumerable<int>

  implements: Collections.IEnumerable

val W : ‘a array

  type: ‘a array

  implements: ICloneable

  implements: Collections.IList

  implements: Collections.ICollection

  implements: Collections.IStructuralComparable

  implements: Collections.IStructuralEquatable

  implements: Collections.Generic.IList<‘a>

  implements: Collections.Generic.ICollection<‘a>

  implements: seq<‘a>

  implements: Collections.IEnumerable

  inherits: Array

property Array.Length: int
val Munproposed : (int * int list) list

  type: (int * int list) list

  implements: Collections.IStructuralEquatable

  implements: IComparable<List<int * int list>>

  implements: IComparable

  implements: Collections.IStructuralComparable

  implements: Collections.Generic.IEnumerable<int * int list>

  implements: Collections.IEnumerable

Multiple items

module List

from Microsoft.FSharp.Collections

——————–

type List<‘T> =

  | ( [] )

  | ( :: ) of ‘T * ‘T list

  with

    interface Collections.IEnumerable

    interface Collections.Generic.IEnumerable<‘T>

    member Head : ‘T

    member IsEmpty : bool

    member Item : index:int -> ‘T with get

    member Length : int

    member Tail : ‘T list

    static member Cons : head:’T * tail:’T list -> ‘T list

    static member Empty : ‘T list

  end

Full name: Microsoft.FSharp.Collections.List<_>

  type: List<‘T>

  implements: Collections.IStructuralEquatable

  implements: IComparable<List<‘T>>

  implements: IComparable

  implements: Collections.IStructuralComparable

  implements: Collections.Generic.IEnumerable<‘T>

  implements: Collections.IEnumerable

val init : int -> (int -> ‘T) -> ‘T list

Full name: Microsoft.FSharp.Collections.List.init

val M : ‘a array

  type: ‘a array

  implements: ICloneable

  implements: Collections.IList

  implements: Collections.ICollection

  implements: Collections.IStructuralComparable

  implements: Collections.IStructuralEquatable

  implements: Collections.Generic.IList<‘a>

  implements: Collections.Generic.ICollection<‘a>

  implements: seq<‘a>

  implements: Collections.IEnumerable

  inherits: Array

val mi : int

  type: int

  implements: IComparable

  implements: IFormattable

  implements: IConvertible

  implements: IComparable<int>

  implements: IEquatable<int>

  inherits: ValueType

val sortFun : (int -> float)
val wi : int

  type: int

  implements: IComparable

  implements: IFormattable

  implements: IConvertible

  implements: IComparable<int>

  implements: IEquatable<int>

  inherits: ValueType

val sortBy : (‘T -> ‘Key) -> ‘T list -> ‘T list (requires comparison)

Full name: Microsoft.FSharp.Collections.List.sortBy

val findMarriages : ((int * int list) list -> Map<int,(int * int list)> -> Map<int,(int * int list)>)
val mSingle : (int * int list) list

  type: (int * int list) list

  implements: Collections.IStructuralEquatable

  implements: IComparable<List<int * int list>>

  implements: IComparable

  implements: Collections.IStructuralComparable

  implements: Collections.Generic.IEnumerable<int * int list>

  implements: Collections.IEnumerable

val wEngaged : Map<int,(int * int list)>

  type: Map<int,(int * int list)>

  implements: IComparable

  implements: Collections.Generic.IDictionary<int,(int * int list)>

  implements: Collections.Generic.ICollection<Collections.Generic.KeyValuePair<int,(int * int list)>>

  implements: seq<Collections.Generic.KeyValuePair<int,(int * int list)>>

  implements: Collections.IEnumerable

val bachelors : (int * int list) list

  type: (int * int list) list

  implements: Collections.IStructuralEquatable

  implements: IComparable<List<int * int list>>

  implements: IComparable

  implements: Collections.IStructuralComparable

  implements: Collections.Generic.IEnumerable<int * int list>

  implements: Collections.IEnumerable

val rest : int list

  type: int list

  implements: Collections.IStructuralEquatable

  implements: IComparable<List<int>>

  implements: IComparable

  implements: Collections.IStructuralComparable

  implements: Collections.Generic.IEnumerable<int>

  implements: Collections.IEnumerable

val m : int * int list
Multiple items

module Map

from Microsoft.FSharp.Collections

——————–

type Map<‘Key,’Value (requires comparison)> =

  class

    interface Collections.IEnumerable

    interface IComparable

    interface Collections.Generic.IEnumerable<Collections.Generic.KeyValuePair<‘Key,’Value>>

    interface Collections.Generic.ICollection<Collections.Generic.KeyValuePair<‘Key,’Value>>

    interface Collections.Generic.IDictionary<‘Key,’Value>

    new : elements:seq<‘Key * ‘Value> -> Map<‘Key,’Value>

    member Add : key:’Key * value:’Value -> Map<‘Key,’Value>

    member ContainsKey : key:’Key -> bool

    override Equals : obj -> bool

    member Remove : key:’Key -> Map<‘Key,’Value>

    member TryFind : key:’Key -> ‘Value option

    member Count : int

    member IsEmpty : bool

    member Item : key:’Key -> ‘Value with get

  end

Full name: Microsoft.FSharp.Collections.Map<_,_>

  type: Map<‘Key,’Value>

  implements: IComparable

  implements: Collections.Generic.IDictionary<‘Key,’Value>

  implements: Collections.Generic.ICollection<Collections.Generic.KeyValuePair<‘Key,’Value>>

  implements: seq<Collections.Generic.KeyValuePair<‘Key,’Value>>

  implements: Collections.IEnumerable

val tryFind : ‘Key -> Map<‘Key,’T> -> ‘T option (requires comparison)

Full name: Microsoft.FSharp.Collections.Map.tryFind

union case Option.None: Option<‘T>
val add : ‘Key -> ‘T -> Map<‘Key,’T> -> Map<‘Key,’T> (requires comparison)

Full name: Microsoft.FSharp.Collections.Map.add

union case Option.Some: ‘T -> Option<‘T>
val m’ : int * int list
val mi’ : int

  type: int

  implements: IComparable

  implements: IFormattable

  implements: IConvertible

  implements: IComparable<int>

  implements: IEquatable<int>

  inherits: ValueType

val empty<‘Key,’T (requires comparison)> : Map<‘Key,’T> (requires comparison)

Full name: Microsoft.FSharp.Collections.Map.empty

val map : (‘Key -> ‘T -> ‘U) -> Map<‘Key,’T> -> Map<‘Key,’U> (requires comparison)

Full name: Microsoft.FSharp.Collections.Map.map

val alignJaroWinkler : (‘a array -> ‘a array -> Map<int,int>)

Full name: Snippet.alignJaroWinkler


6
Mar 10

Abstract Thoughts about F# Abstractions

My recent work on Professional F# 2.0 has left me thinking a lot about the nature of abstractions.

In computation, different types of abstraction are trade-offs between clarity and freedom in different areas.   As you move away from the machine architecture, you lose the freedom to push bits around in whichever way you might want.  In the case of garbage collection, this means you might not be able to hand optimize your memory usage.  For managed code, this might mean you can no longer hand optimize your code to a CPU-instruction level.  With even a small degree of abstraction, your choices are cut down to only a small subset of what was previously possible.

In exchange for this loss, you get to think about problems in a constrained way.  This means you can make more valid assumptions about what is going on.  You can express ideas cleanly, without additionally expressing the machine’s management of the resources involved in those ideas.  You can manage less and express more.

The benefit of these constraints goes beyond just what you personally have to worry about while programming.  Some types of abstraction, such as function signature standardization or managed assemblies, enable post-compilation binary compatibility.  Other types of abstraction, such as immutability, allow the compiler to do high level optimizations beyond what was previously possible.

So, while you might be able to do everything with the JMP instruction that you can with C#’s foreach, it takes much more effort to express the same idea with JMP, and a lot more can go wrong.  Similarly, while you may be able to express all of the ideas in C# that you can in F#, it takes much more effort to express those same ideas in C#, and a lot more can go wrong.

And who knows… In the future, we F# users might even get some sweet optimizations out of the deal.