Multiple alignment is a powerful method for identifying and visualizing conserved regions in family of homologous proteins.  Moreover, all phylogenetic methods need multiple alignment to start with. The 'correctness' of this alignment determines the 'correctness' of resulting phylogenetic tree.
 

Alignment methods:

 Different approaches:

* Maximizing (or minimizing) score.  What score?
Find a good function and try to align sequences in order to optimize this function.
MSA
SAGA

* Progressive pairwise or others
Align most similar pairs with each other,then merge and/or optimize those alignments.
Some of them start with finding local similarities in aligned sequences (local-global approach).

Global	Local-global
ClustalX  (Thompson,1994) ClustalW (Thompson,1994) PileUp MAP	DIALIGN (Morgenstern, 1996) DFALIGN Macaw (Schuler,1991) IterAlign (Brocchieri,1998) PIMA prrp (Gotoh, 1996)

* Probabilistic approach (Profiles and Hidden Markov Models)
Find model, align each sequence to model independently.
ClustalW - does simple profile alignment
MaxHom - iterative search of database with profiles (Sander,1991)
HMMER 2.0 package (HMMbuild + HMMalign)  by Sean Eddy et al., WashU
COVE for RNA alignment  by Sean Eddy and Richard Durbin, WashU (Eddy,1994)
SAM  by University of California, Santa Cruz

Detailed description for each method follows:

Possible scoring functions for multiple alignments:

The scoring matrices and gap costs that are used in pairwise alignments cannot be directly and easily applied to multiple alignments. The methods that are dependent on optimal score should first find out how to estimate this score of 'goodness'.
Ideally we would like that best alignment would have highest score. But what is the best alignment? What criterion can be used for comparing the 'goodness' of alignments? Most of the recent papers that compare multiple alignment methods use structural alignments as "criterion of truth".  These are based on protein families whose crystal structure has been established and that can be aligned based on their secondary structure regions (See HSSP database or FSSP database for examples). The score is then given as percentage of residues that are correctly aligned in structurally conserved regions.
The classical way for calculating multiple alignment score is to sum all possible  pairwise scores. The main problem is that this system does not guarantee that the best score is reflecting biologically best alignment.
An alternative method for scoring multiple alignments was published recently: the COFFEE method. This method tries to achieve maximum agreement with all pairwise alignments. Thus it is fairly similar to the classical sum of pairs method. The COFFEE is reported to result in slightly better alignments when sequences are only moderately related (<40% identical).
 

Optimization methods:

The idea for using optimization methods for multiple alignment is derived from analogy with pairwise alignment: all pairwise alignment methods are based on optimization of alignment score.  It would be pretty natural to transfer the same logic to multiple alignment. However, the dynamic programming algorithm gets too complex when used on more than 2 sequences. The main concern here is computer memory - it needs memory in order of N^k where N is average length and k is number of sequences. In practice this means that for aligning of 6 average protein sequences you would need more than million of Megabytes of memory. The approximation algorithms (like FASTA in pairwise alignments) can be used, but they still cannot handle more than 8 sequences with maximum length of 100.  The MSA program is the best example of such efforts (available from BCM, Texas).  It still needs up to 100 Mb memory if you run it on your own computer.
So why should we use those memory-hungry methods? Well, keep in mind that dynamic programming is guaranteed to return the best alignment (with given scoring matrix and gap cost). So it is worth trying. MSA still gives you this best alignment, with little tweaking of parameters (parameter epsilon determines the scope of search and is giving you compromise between search speed and probability of getting best alignment).
A different approach to optimization method has been taken recently. The idea is that we can find the best score without trying all possibilities. This sounds illogical, but is nearly true. The random testing methods have widespread use in solving complicated problems like this. More specifically, GENETIC ALGORITHM was used for finding best alignment in program SAGA. With genetic algoritms, the multiple alignment is changed and improved in evolution-like manner, eventually evolving to (hopefully) best scoring alignment. It takes some time (1 hour for 6 proteins of 330 amino-acids on RS10000 processor), but  is still much more affordable, mainly because of  tiny memory requirements.

All optimization methods are dependent on correct choice of  gap costs and scoring matrices.

Progressive alignment methods:

Global alignment methods

• similarity of sequences (measured from guide tree)

• scoring matrix used
• similarity of sequences
• context (hydrophilicity, existence of gaps in other sequences)
• length of sequences (Gap opening penalty)
• difference in sequence length (Gap extension penalty)

Local alignment methods

From practical side, I would suggest to try some local multiple alignment programs and see how do you like them. The outcome will depend on your sequences and there is no good criterions to say one is better than others so you have to try.
You may want to test:
MacaW (Windows and Macintosh), that is statistically well grounded and allows editing of alignment,
Iteralign, that has the WWW-based service,
prrp by Osamu Gotoh. The latter is UNIX program that uses iterations to make alignment, phylogenetic tree and sequence weights mutually consistent. Because of optimizing iterations, it could also be classified as optimizing alignment method. Claimed to be good but is not simple to use.

More complicated methods code the information of old alignment into the profile and can use such profiles for alignment. Basically profiles are probabilities of finding given amino acid at given position. Using profile for scoring alignment is technically almost as easy as using usual scoring matrices for scoring each alignment position. Usual scoring matrices also reflect probabilities of one amino acid changing to another. But PAM, BLOSUM and Gonnet matrices contain generalized information from all proteins. The main advantage of profiles is that they specify these probabilities for each amino acid position of protein separately. Of course they are not as general as scoring matrices, you need to create different profile for each family of proteins. But the gain in alignment quality is well worth the effort.  A good example of intelligent use of profiles for sequence alignment is PSI-BLAST program, that creates alignment and profile to extract more members of given family from database.

Even more advanced methods are Hidden Markov Model (HMM) based methods. They use a sequence model to store probabilities of amino acid substitutions and gap frequencies.  They are similar to profile methods, except that their model can contain more information than a simple profile.  Most often the gap penalty values are coded into the HMM model, thus eliminating the need for user defined gap penalties, the source of many if not most errors in multiple alignments. The most used program for building HMM models and aligning sequences according to HMM model is HMMer. The collection of thousands of existing HMM models can be found from Pfam database.
Beside the well-grounded gap-penalty handling, HMM's have another advantage in multiple alignments. The speed of alignment. The speed of alignment is in linear order with number of sequences as opposed to exponential order for most other alignment methods. Thus, HMMs are particularly useful for protein families with many members.

HMM models are not limited to profile and gap modelling. For aligning RNA molecules, the incorporation of secondary structure information into the HMM model is possible. Nevertheless, the number of calculations increases with additional information, thus this program COVE currently does not handle RNA sequences longer than few hundred nucleotides.