Held: Wednesday, 10 November 2004
Summary:
Today we consider one of the most important implementations of
dictionaries, Hash tables.
Related Pages:
Overview:
- The Idea of Hashing.
- Hash Functions.
- Hashing in Java.
- Handling Hashing Conflicts.
- As you may have noticed, although dictionaries are
a lot like arrays
(except that you index dictionaries by object and arrays by number).
- However, in terms of implementation, that claim hasn't held.
- The unordered list implementation has O(n) set and
O(n) get.
- The ordered array implementation has O(n) set and
O(log2n) get.
- The binary search tree implementation has O(logn) set and
O(logn) get, if we're lucky.
- The corresponding array operations are O(1).
- Is there a dictionary implementation with O(1)
get
and set
?
- Surprisingly, if you're willing to sacrifice some space and increase
your constant, it is possible to build a dictionary that is likely
to have O(1)
get
and set
.
- How? Well, we know that arrays provide O(1) get and set, so use arrays.
- How do we use an array? We "number" the keys in such a way that
- all numbers are between 0 and array.length-1
- no two keys have the same number (or at least few have the same
number).
- If there are no collisions (keys with the same number), the system is simple
- To set a value, determine the number corresponding to the key
and add it in that place of the array. This is O(1+cost of
computing that number).
- To get a value, determine the number corresponding to the key
and look in the appropriate cell. This is O(1+cost of finding that
number).
- Implementations of dictionaries using this strategy are called
hash tables.
- The function used to convert an object to a number is the
hash function.
- To better understand hash tables, we need to consider
- The hash functions we might develop.
- What to do about collisions (when two keys map to the same index).
- The goal in developing a hash function is to come up with a function
that is unlikely to map two objects to the same position.
- Now, this isn't possible (particularly if we have more objects than
positions).
- We'll discuss what to do about two objects mapping to
the same position later.
- Hence, we sometimes accept a situation in which the hash function
distributes the objects more or less uniformly.
- It is worth some experimentation to come up with such a function.
- In addition, we should consider the cost of computing the hash function.
We'd like something that is relatively low cost (not just constant time,
but not too many steps within that constant).
- We'd also like a function that does (or can) give us a relatively
large range of numbers, so that we can get fewer collisions by increasing
the size of the hash table.
- We might want to make the size of the table a parameter to the
hash function.
- We might strive for a hash function that uses the range of positive
integers, and mod it by the size of the table.
- What are some hash functions you might use for strings?
- Sum the ASCII values in the string
- constant1*firstLetter + constant2*secondLetter
- ...
- Hash tables are so useful that Java includes them as a standard
library class,
java.util.Hashtable
.
- Let's look over the
- documentation for
java.util.Hashtable
.
- Why are there three constructors?
- What methods are there other than
get
and set
?
- Where's the hash function?
- Answer: Every class is expected to provide its own.
- Since we typically have many more possible keys than places in the
array, we will eventually hit a point in which two keys hash to the
same value.
- What should we do when this conflict happens?
- It is not reasonable to disallow the second key, since it's a
different key.
- We could put a list of key/value pairs, rather than a singleton, at
every position.
- We could step through the array until we find an empty space.
- Our analysis of Hash Tables to date has been based on two simple
operations: get and set.
- What happens if we want to remove elements? This operation
can significantly complicate matters.
- If we've chosen the
shift into a blank space
technique for
resolving collisions, what do we
do when it comes time to remove elements?
- Do we shift everything back? If so, think about how far we may have
to look.
- Do we leave the thing there as a blank? We might then then remove
it later when it's convenient to do so.
- Do we do something totally different?
- Note also that there are different ways of specifying
remove
. We
might remove the element with a particular key. We might instead remove
elements based on their value. The second is obviously a much slower
operation than the first (unless we've developed a special way to handle
that problem - see if you can think of one).