TL;DR: There are so many addresses that it is improbable that anyone will ever generate a duplicate of another address in use – as long as random number generators work as they should.
2^160 possible addresses
Bitcoin addresses consist of an alphanumerical string with a length of up to 34 characters, excluding the capital "O", the capital "I" and the lowercase "l", as well as the number "0". This would allow for 58^34 possible combinations, however, as some of the positions are used for a checksum this is reduced to 2^160 valid addresses. The checksum on the other hand allows to detect mistyped addresses as invalid, so that it is highly unlikely to accidentally input another valid address.
Chance of 3.42*10^(-27) for a collision at one trillion addresses
Even if we generously assume that at some point there will be one trillion addresses (approx. 160 for each of this planet's population), according to the simple approximation formula for the Birthday attack given on Wikipedia, the chance is 3.42*10^(-27) that any two of those Bitcoin addresses collide. In comparison, the figure of addresses that were ever used to receive bitcoins was just over 13 million in May 2013.
Address space could be augmented
Especially as the protocol at any point could be adapted to accept even longer addresses, we can say, yes, it is theoretically possible, but unlikely enough that we can assume for our purposes that it will never happen.*
It's impractical to generate addresses for gain
To answer your other question: For each address there apparently are 2^96 different private keys whose corresponding public key will map to the same address. So to actually try to pursue Bitcoin theft, it would be much more sensible to just generate random private keys and hope to find one that has a corresponding address with money on it. In all likelihood the power for such calculations would cost more than anyone could earn with such a scheme.
*As long as certain random number generators work as well as they should... :)