Introduction
I think most of the existing hash functions were developed many years ago, by very smart people. It is probably not possible to improve hashing algorithms any longer, still, I had to try this idea below.
Background
But first, as you know, a hash function is a function that converts any data of any size to a fixed sized number. It is widely used in, e.g., lookup tables.
In a previous article of mine, on chess engine programming1, I learned a very cool hashing technique for chess board states. Zobrist hashing5 proved to be able to give an almost unique hash code for every chess position in a game, even though there are almost infinitive states a chess game can be in. I think that we can use Zobrist algorithm to hash for example something more common like words. Results revealed at the end. (It worked!)
As you probably know, .NET already contains a hash function for string
s. The method is called GetHashCode
and it converts any object to an integer.
I want to know if we can create a better function using Zobrist ideas and with less collisions.
Hash Collisions
A hash collision occurs when two different objects, in this case string
s, result in the same hash code.
I'm using a list of 466,545 English words3. The function GetHashCode
in .NET produces only 48 collisions on this data set, so it is probably good enough in most scenarios. If you are fine with 48 collisions, you don’t have to read on. Here are few examples of words with same .NET hash code.
- furnaces ferrate
- matinee cistronic
- toeholds argon
- unsmart pulped
You might argue that since the dataset is known, it would be very easy to create a perfect hash algorithm without collisions. We could just give each word its unique number.
However, that is not the purpose of this investigation. We want the hash function to work for any word in any language and with a very low probability of collisions.
Zobrist Hash
Zobrist hashing works by bit-wise XOR-ing random numbers. In a chess engine, random numbers are generated for moves, and a move is defined by removing a piece from a square and then putting it on another.
There are two player colors, six piece types and 64 squares. 2 * 6 * 64 = 768 random numbers.
(Not taking into account castling and enpassant, but you get the point.)
When hashing words, I thought that we can generate random numbers for each character and then do the same thing as in Zobrist hashing. One bit-wise XOR operation for every character in a word.
public static int ZobHash(this string text)
{
var hash = ZobRandom.Start;
foreach (var c in text)
hash ^= ZobRandom.Numbers[c];
return hash;
}
But when doing this, I get 218,163 collisions. Not very encouraging, but instead of giving up, let’s look at why there are so many collisions.
Here are a few examples:
- abied abide
- abiegh abeigh
- acidly acidyl
- acron acorn
As you can see, Zobrist hash gives the same hash code for words independent of the order of the letters, i.e., the order of the XOR operations.
That is how XOR works and it's perfect for storing chess game states, because the order of moves that led to a board position doesn’t matter, it is still the same game state. But for hashing of words, it does not work.
Improve Zobrist Algorithm
Let’s try to extend Zobrist hash so the order of the letters also affects the resulting hash code.
What if we try to do a cyclic bit-shift (very fast operations) for each character random number?
A cyclic bit shift pushes the bits in a number to the left (or to the right if you want) like this:
When the number 153, which is binary | 10011001 |
is cyclically shifted to the left, the result is | 00110011 |
and when you push it two times (next char) | 01100110 |
This is the function for cyclic left bit shift.
private static uint RotateLeft(uint value, int count)
{
var r = count % 32;
return (value << r) | (value >> (32 - r));
}
And the final hash function looks like:
public static uint ZobHash(this string text)
{
var hash = ZobRandom.Start;
var i = 1;
foreach (var c in text)
hash ^= RotateLeft(ZobRandom.Numbers[c], i++);
return hash;
}
Random Seeds
As I mentioned above, Zobrist hash uses a list of random numbers. For some reason, the function works better with some random numbers.
You might be unlucky to select a bad random seed that in the end generates more hash collisions.
So, implementing a Zobrist hashing function is also about selecting a good random seed.
The Results
Below is the result of counting performance and number of collisions for the list of 466,545 unique English words.
Also comparing with a few other hash functions from Brandon Dahler nuget System.Data.HashFunction4.
If you have a suggestion on more hash functions we should compare with, please comment below.
These are the results from hashing byte-arrays.
Functions with 32-bit hash codes:
Function | 106 Words/s | Collisions |
ZobHash | 4.67 | 5 |
GetHashCode .net | 7.52 | 1592 |
JenkinsOneAtATime | 3.99 | 36 |
BernsteinHash | 4.57 | 343 |
CRC | 3.17 | 23 |
ELF64 | 4.32 | 4347 |
JenkinsLookup2 | 3.67 | 23 |
JenkinsLookup3 | 3.64 | 30 |
ModifiedBernsteinHash | 5.69 | 320 |
MurmurHash1 | 3.86 | 27 |
xxHash | 3.15 | 16 |
Djb2 | 5.55 | 344 |
Functions with 64-bit hash codes:
Function | 106 Words/s | Collisions |
ZobHashLong | 5.3 | 0 |
FNV1 | 3.24 | 0 |
FNV1a | 3.11 | 0 |
These are the results from hashing string
s (a lot faster without conversion to byte-array):
Hash function | 106 Words/s | Collisions |
ZobHash | 24.6 | 5 |
ZobHashLong | 25.9 | 0 |
GetHashCode .net | 51.8 | 48 |
Conclusion
ZobHash performs well compared to all tested functions. For functions that generate 32-bit hash codes, it has the lowest rate of collisions. There are functions that are faster, but they produce more collisions.
Interestingly .NET GetHashCode
gives different number of collisions depending on if you hash a byte array or a string
.
For 64-bit hash code functions, none of the tested functions generates collisions but ZobHash is the fastest.
Links
- Test-Driven-Chess-Artificial-Intelligence
- https://en.wikipedia.org/wiki/Hash_function
- https://github.com/dwyl/english-words
- https://github.com/brandondahler/Data.HashFunction/
- https://en.wikipedia.org/wiki/Zobrist_hashing
That was all for now and don't forget to vote!
Thanks!
History
- January 16, 2018 - Version 1.0.0
- January 19, 2018 - Version 1.0.1
- Fixed a bug on right shift for
int
s. Changed int
to uint
- Added simple implementation in C
This member has not yet provided a Biography. Assume it's interesting and varied, and probably something to do with programming.