Finding a good hash function can be difficult. Most of the common
methods (polynomial multipliers, shifitng bits, mid-square techniques)
don't seem to change the number of collisions significantly on the
file /usr/dict/words.

A technique that can be used is a STOCHASTIC LEARNING algorithm. One variable that
will change the number of collisions is to change the order in which
each letter of each word is used in a polynomial hash sequence. For
example, the word "cat" can be hashed as:

c + a*39 + t*39^2

OR

a + t*39 + c*39^2

OR

t + a*39 + c*39^2

etc. Essentially this is assigning an exponent to be associated with a
character position in the word..In the example above, there are
exactly 3! = 3 x 2 x 1 = 6 sequences of exponents associated with a 3
character word. 

So we can try different sequences on a set of words to be hashed.
For the Java reserved words, we have a maximum word length of 12 ("synchronized").
So there are 12! possible exponent sequences to test.

PROBLEM: this is an enormous number of sequences!

its equal to: (12!=479,001,600). This is probably too many
for a brute force trial and error approach.

So what do we do?  We try RANDOM sequences and
see how many collisions we get. If we get a reduced number of
collisions we remember that sequence, and keep generating new
sequences.  So the algorithm competes against itself, remembering the
best sequence over time.