Decision Tree learning algorithm - algorithm

I want to preface this by saying that this is a homework assignment.
I am given a set of Q binary input variables that will be used to classify output of Y which is also binary.
The first part of the question is: at most how many examples do I need to enumarate all possibile combinations of Q? I am currently think that since it asks for at most I will need Q as it is possible that all values up to Q-1 are the same for instance 1 and the item at Q is 0 .
The second part of the question is: at most how many leaf nodes can the tree have given Z examples?
My current answer is that at most the tree would have 2 leaf nodes, one representing true and one representing false since it is dealing with binary inputs and binary outputs.
Is this the correct way of examining this problem or am I generalizing my answers too deeply?
Edit
After looking at Cameron's response, I would now turn my first answer into 2^Q and to build on his example of Q = 3, I would get 2^3 or 8 (2*2*2). Please correct if that is incorrect thinking.
Edit #2
The second part of the question it appears as though it should be (2^Q) * Z or to provide an example: (2^3) * 3) or 8*3 = 24 leaf nodes. To recap if I have 3 inputs that are binary I would initially take 2^3 and get 8 now I want to go over 3 examples. Therefore I should get 8*3 or 24.
Edit #3
In hindsight it seems that no matter how many examples I use the number of leaf nodes should never increase, as it is a per tree basis.

I'd suggest you approach the problem by working out small example cases by hand.
For the first part, choose a small value for Q, say 3, and write down all possible combinations of Q. Then you can figure out how many examples you need. Increase Q and do it again.
For the second part of your question, pick a small Z and run the decision tree algorithm by hand. See how many leaves you get. Then pick another Z and see if/how it changes. Try generating different examples (with the same Z) and see if you can change the number of leaves.

Related

Efficient algorithm for eliminating nodes in "graph"?

Suppose I have a a graph with 2^N - 1 nodes, numbered 1 to 2^N - 1. Node i "depends on" node j if all the bits in the binary representation of j that are 1, are also 1 in the binary representation of i. So, for instance, if N=3, then node 7 depends on all other nodes. Node 6 depends on nodes 4 and 2.
The problem is eliminating nodes. I can eliminate a node if no other nodes depend on it. No nodes depend on 7; so I can eliminate 7. After eliminating 7, I can eliminate 6, 5, and 3, etc. What I'd like is to find an efficient algorithm for listing all the possible unique elimination paths. (that is, 7-6-5 is the same as 7-5-6, so we only need to list one of the two). I have a dumb algorithm already, but I think there must be a better way.
I have three related questions:
Does this problem have a general name?
What's the best way to solve it?
Is there a general formula for the number of unique elimination paths?
Edit: I should note that a node cannot depend on itself, by definition.
Edit2: Let S = {s_1, s_2, s_3,...,s_m} be the set of all m valid elimination paths. s_i and s_j are "equivalent" (for my purposes) iff the two eliminations s_i and s_j would lead to the same graph after elimination. I suppose to be clearer I could say that what I want is the set of all unique graphs resulting from valid elimination steps.
Edit3: Note that elimination paths may be different lengths. For N=2, the 5 valid elimination paths are (),(3),(3,2),(3,1),(3,2,1). For N=3, there are 19 unique paths.
Edit4: Re: my application - the application is in statistics. Given N factors, there are 2^N - 1 possible terms in statistical model (see http://en.wikipedia.org/wiki/Analysis_of_variance#ANOVA_for_multiple_factors) that can contain the main effects (the factors alone) and various (2,3,... way) interactions between the factors. But an interaction can only be present in a model if all sub-interactions (or main effects) are present. For three factors a, b, and c, for example, the 3 way interaction a:b:c can only be in present if all the constituent two-way interactions (a:b, a:c, b:c) are present (and likewise for the two-ways). Thus, the model a + b + c + a:b + a:b:c would not be allowed. I'm looking for a quick way to generate all valid models.
It seems easier to think about this in terms of sets: you are looking for families of subsets of {1, ..., N} such that for each set in the family also all its subsets are present. Each such family is determined by the inclusion-wise maximal sets, which must be overlapping. Families of pairwise overlapping sets are called Sperner families. So you are looking for Sperner families, plus the union of all the subsets in the family. Possibly known algorithms for enumerating Sperner families or antichains in general are useful; without knowing what you actually want to do with them, it's hard to tell.
Thanks to #FalkHüffner's answer, I saw that what I wanted to do was equivalent to finding monotonic Boolean functions for N arguments. If you look at the figure on the Wikipedia page for Dedekind numbers (http://en.wikipedia.org/wiki/Dedekind_number) the figure expresses the problem graphically. There is an algorithm for generating monotonic Boolean functions (http://www.mathpages.com/home/kmath094.htm) and it is quite simple to construct.
For my purposes, I use the algorithm, then eliminate the first column and last row of the resulting binary arrays. Starting from the top row down, each row has a 1 in the ith column if one can eliminate the ith node.
Thanks!
You can build a "heap", in which at depth X are all the nodes with X zeros in their binary representation.
Then, starting from the bottom layer, connect each item to a random parent at the layer above, until you get a single-component graph.
Note that this graph is a tree, i.e., each node except for the root has exactly one parent.
Then, traverse the tree (starting from the root) and count the total number of paths in it.
UPDATE:
The method above is bad, because you cannot just pick a random parent for a given item - you have a limited number of items from which you can pick a "legal" parent... But I'm leaving this method here for other people to give their opinion (perhaps it is not "that bad").
In any case, why don't you take your graph, extract a spanning-tree (you can use Prim algorithm or Kruskal algorithm for finding a minimal-spanning-tree), and then count the number of paths in it?

Algorithm determining the smallest coprime subset

Given a set A of n positive integers, determine a non-empty subset B
consisting of as few elements as possible such that their GCD is 1 and output its size.
For example: 5 6 10 12 15 18
yields an output of "3", while:
5 2 4 6 8 10
equals "NONE" since no subset can be determined.
So it seems really basic but I'm still stuck with it. My thoughts on it are as follows: we know that having the multiples of some number already present in the set are useless since their divisors are the same times some factor k and we're going for the smallest subsest. Hence, for every ni, we remove any kni where k is a positive int from further calculations.
That's where I get stuck, though. What should I do next? I can only think of a dumb, brute force approach of trying if there is already some 2-element subset, then 3-elem and so on. What should I check to determine it in some more clever way?
Suppose for each A,B (two elements) we calculate their greatest common
divisor D. And then we store these D values somewhere as a map of the form:
A,B -> D
Let's say we also store the reverse map
D -> A,B
If there's at least one D=1 then there we go - the answer is 2.
Suppose now, there's no such D that D=1.
What condition should be met for the answer to be 3?
I think this one:
there exist two D values say D1 and D2 such that GCD(D1, D2)=1.
Right?
So now instead of As and Bs, we've transformed our problem to the
same problem over the set of all Ds and we've transformed the option of
a the 2 answer to the option a 3 answer. Right?
I am not 100% sure just thinking out loud.
But this transformed problem is even worse as
we have to store much more values.
(combinations of N elements class 2).
Not sure, this problem you pose seems like a hard
problem to me. I would be surprised if there exists
a better approach than brute-force
and would be interested to know it.
What you need to think on (and look for) is this:
is there a way to express GCD(a1, a2, ... aN)
if you know their pair-wise GCDs. If there's some
sort of method or formula you can simplify a bit
your search (for the smallest subset matching
the desired criterion).
See also this link. Maybe it could help.
https://cs.stackexchange.com/questions/10249/finding-the-size-of-the-smallest-subset-with-gcd-1
The problem is definitely a tough one to solve. I can't see any computationally efficient algorithm that would guaranteed find the solution in reasonable time.
One approach is:
Form a list of ordered sets that would contain the prime factors of each element in the original set.
Now you need to find the minimum number of sets for which their intersection is zero.
To do that, first order these sets in your list so that the sets that have least number of intersections with other sets are towards the beginning. Now what are "least number of intersections"?
This is where heuristics come into play. It can be:
1. set having Less of MIN number of intersections with other elements.
2. set having Less of MAX number of intersections with other elements.
3. Any other more suitable definition.
Now you will need to expensively iterate through all the combinations maybe through recursion to determine the solution.

Algorithm for finding basis of a set of bitstrings?

This is for a diff utility I'm writing in C++.
I have a list of n character-sets {"a", "abc", "abcde", "bcd", "de"} (taken from an alphabet of k=5 different letters). I need a way to observe that the entire list can be constructed by disjunctions of the character-sets {"a", "bc", "d", "e"}. That is, "b" and "c" are linearly dependent, and every other pair of letters is independent.
In the bit-twiddling version, the character-sets above are represented as {10000, 11100, 11111, 01110, 00011}, and I need a way to observe that they can all be constructed by ORing together bitstrings from the smaller set {10000, 01100, 00010, 00001}.
In other words, I believe I'm looking for a "discrete basis" of a set of n different bit-vectors in {0,1}k. This paper claims the general problem is NP-complete... but luckily I'm only looking for a solution to small cases (k < 32).
I can think of really stupid algorithms for generating the basis. For example: For each of the k2 pairs of letters, try to demonstrate (by an O(n) search) that they're dependent. But I really feel like there's an efficient bit-twiddling algorithm that I just haven't stumbled upon yet. Does anyone know it?
EDIT: I ended up not really needing a solution to this problem after all. But I'd still like to know if there is a simple bit-twiddling solution.
I'm thinking a disjoint set data structure, like union find turned on it's head (rather than combining nodes, we split them).
Algorithm:
Create an array main where you assign all the positions to the same group, then:
for each bitstring curr
for each position i
if (curr[i] == 1)
// max of main can be stored for constant time access
main[i] += max of main from previous iteration
Then all the distinct numbers in main are your different sets (possibly using the actual union-find algorithm).
Example:
So, main = 22222. (I won't use 1 as groups to reduce possible confusion, as curr uses bitstrings).
curr = 10000
main = 42222 // first bit (=2) += max (=2)
curr = 11100
main = 86622 // first 3 bits (=422) += max (=4)
curr = 11111
main = 16-14-14-10-10
curr = 01110
main = 16-30-30-26-10
curr = 00011
main = 16-30-30-56-40
Then split by distinct numbers:
{10000, 01100, 00010, 00001}
Improvement:
To reduce the speed at which main increases, we can replace
main[i] += max of main from previous iteration
with
main[i] += 1 + (max - min) of main from previous iteration
EDIT: Edit based on j_random_hacker's comment
You could combine the passes of the stupid algorithm at the cost of space.
Make a bit vector called violations that is (k - 1) k / 2 bits long (so, 496 for k = 32.) Take a single pass over character sets. For each, and for each pair of letters, look for violations (i.e. XOR the bits for those letters, OR the result into the corresponding position in violations.) When you're done, negate and read off what's left.
You could give Principal Component Analysis a try. There are some flavors of PCA designed for binary or more generally for categorical data.
Since someone showed it as NP complete, for large vocabs I doubt you will do better than a brute force search (with various pruning possible) of the entire set of possibilities O((2k-1) * n). At least in a worst case scenario, probably some heuristics will help in many cases as outlined in the paper you linked. This is your "stupid" approach generalized to all possible basis strings instead of just basis of length 2.
However, for small vocabs, I think an approach like this would do a lot better:
Are your words disjoint? If so, you are done (simple case of independent words like "abc" and "def")
Perform bitwise and on each possible pair of words. This gives you an initial set of candidate basis strings.
Goto step 1, but instead of using the original words, use the current basis candidate strings
Afterwards you also need to include any individual letter which is not a subset of one of the final accepted candidates. Maybe some other minor bookeeping for things like unused letters (using something like a bitwise or on all possible words).
Considering your simple example:
First pass gives you a, abc, bc, bcd, de, d
Second pass gives you a, bc, d
Bookkeeping gives you a, bc, d, e
I don't have a proof that this is right but I think intuitively it is at least in the right direction. The advantage lies in using the words instead of the brute force's approach of using possible candidates. With a large enough set of words, this approach would become terrible, but for vocabularies up to say a few hundred or maybe even a few thousand I bet it would be pretty quick. The nice thing is that it will still work even for a huge value of k.
If you like the answer and bounty it I'd be happy to try to solve in 20 lines of code :) and come up with a more convincing proof. Seems very doable to me.

How do I calculate the shanten number in mahjong?

This is a followup to my earlier question about deciding if a hand is ready.
Knowledge of mahjong rules would be excellent, but a poker- or romme-based background is also sufficient to understand this question.
In Mahjong 14 tiles (tiles are like
cards in Poker) are arranged to 4 sets
and a pair. A straight ("123") always
uses exactly 3 tiles, not more and not
less. A set of the same kind ("111")
consists of exactly 3 tiles, too. This
leads to a sum of 3 * 4 + 2 = 14
tiles.
There are various exceptions like Kan
or Thirteen Orphans that are not
relevant here. Colors and value ranges
(1-9) are also not important for the
algorithm.
A hand consists of 13 tiles, every time it's our turn we get to pick a new tile and have to discard any tile so we stay on 13 tiles - except if we can win using the newly picked tile.
A hand that can be arranged to form 4 sets and a pair is "ready". A hand that requires only 1 tile to be exchanged is said to be "tenpai", or "1 from ready". Any other hand has a shanten-number which expresses how many tiles need to be exchanged to be in tenpai. So a hand with a shanten number of 1 needs 1 tile to be tenpai (and 2 tiles to be ready, accordingly). A hand with a shanten number of 5 needs 5 tiles to be tenpai and so on.
I'm trying to calculate the shanten number of a hand. After googling around for hours and reading multiple articles and papers on this topic, this seems to be an unsolved problem (except for the brute force approach). The closest algorithm I could find relied on chance, i.e. it was not able to detect the correct shanten number 100% of the time.
Rules
I'll explain a bit on the actual rules (simplified) and then my idea how to tackle this task. In mahjong, there are 4 colors, 3 normal ones like in card games (ace, heart, ...) that are called "man", "pin" and "sou". These colors run from 1 to 9 each and can be used to form straights as well as groups of the same kind. The forth color is called "honors" and can be used for groups of the same kind only, but not for straights. The seven honors will be called "E, S, W, N, R, G, B".
Let's look at an example of a tenpai hand: 2p, 3p, 3p, 3p, 3p, 4p, 5m, 5m, 5m, W, W, W, E. Next we pick an E. This is a complete mahjong hand (ready) and consists of a 2-4 pin street (remember, pins can be used for straights), a 3 pin triple, a 5 man triple, a W triple and an E pair.
Changing our original hand slightly to 2p, 2p, 3p, 3p, 3p, 4p, 5m, 5m, 5m, W, W, W, E, we got a hand in 1-shanten, i.e. it requires an additional tile to be tenpai. In this case, exchanging a 2p for an 3p brings us back to tenpai so by drawing a 3p and an E we win.
1p, 1p, 5p, 5p, 9p, 9p, E, E, E, S, S, W, W is a hand in 2-shanten. There is 1 completed triplet and 5 pairs. We need one pair in the end, so once we pick one of 1p, 5p, 9p, S or W we need to discard one of the other pairs. Example: We pick a 1 pin and discard an W. The hand is in 1-shanten now and looks like this: 1p, 1p, 1p, 5p, 5p, 9p, 9p, E, E, E, S, S, W. Next, we wait for either an 5p, 9p or S. Assuming we pick a 5p and discard the leftover W, we get this: 1p, 1p, 1p, 5p, 5p, 5p, 9p, 9p, E, E, E, S, S. This hand is in tenpai in can complete on either a 9 pin or an S.
To avoid drawing this text in length even more, you can read up on more example at wikipedia or using one of the various search results at google. All of them are a bit more technical though, so I hope the above description suffices.
Algorithm
As stated, I'd like to calculate the shanten number of a hand. My idea was to split the tiles into 4 groups according to their color. Next, all tiles are sorted into sets within their respective groups to we end up with either triplets, pairs or single tiles in the honor group or, additionally, streights in the 3 normal groups. Completed sets are ignored. Pairs are counted, the final number is decremented (we need 1 pair in the end). Single tiles are added to this number. Finally, we divide the number by 2 (since every time we pick a good tile that brings us closer to tenpai, we can get rid of another unwanted tile).
However, I can not prove that this algorithm is correct, and I also have trouble incorporating straights for difficult groups that contain many tiles in a close range. Every kind of idea is appreciated. I'm developing in .NET, but pseudo code or any readable language is welcome, too.
I've thought about this problem a bit more. To see the final results, skip over to the last section.
First idea: Brute Force Approach
First of all, I wrote a brute force approach. It was able to identify 3-shanten within a minute, but it was not very reliable (sometimes too a lot longer, and enumerating the whole space is impossible even for just 3-shanten).
Improvement of Brute Force Approach
One thing that came to mind was to add some intelligence to the brute force approach. The naive way is to add any of the remaining tiles, see if it produced Mahjong, and if not try the next recursively until it was found. Assuming there are about 30 different tiles left and the maximum depth is 6 (I'm not sure if a 7+-shanten hand is even possible [Edit: according to the formula developed later, the maximum possible shanten number is (13-1)*2/3 = 8]), we get (13*30)^6 possibilities, which is large (10^15 range).
However, there is no need to put every leftover tile in every position in your hand. Since every color has to be complete in itself, we can add tiles to the respective color groups and note down if the group is complete in itself. Details like having exactly 1 pair overall are not difficult to add. This way, there are max around (13*9)^6 possibilities, that is around 10^12 and more feasible.
A better solution: Modification of the existing Mahjong Checker
My next idea was to use the code I wrote early to test for Mahjong and modify it in two ways:
don't stop when an invalid hand is found but note down a missing tile
if there are multiple possible ways to use a tile, try out all of them
This should be the optimal idea, and with some heuristic added it should be the optimal algorithm. However, I found it quite difficult to implement - it is definitely possible though. I'd prefer an easier to write and maintain solution first.
An advanced approach using domain knowledge
Talking to a more experienced player, it appears there are some laws that can be used. For instance, a set of 3 tiles does never need to be broken up, as that would never decrease the shanten number. It may, however, be used in different ways (say, either for a 111 or a 123 combination).
Enumerate all possible 3-set and create a new simulation for each of them. Remove the 3-set. Now create all 2-set in the resulting hand and simulate for every tile that improves them to a 3-set. At the same time, simulate for any of the 1-sets being removed. Keep doing this until all 3- and 2-sets are gone. There should be a 1-set (that is, a single tile) be left in the end.
Learnings from implementation and final algorithm
I implemented the above algorithm. For easier understanding I wrote it down in pseudocode:
Remove completed 3-sets
If removed, return (i.e. do not simulate NOT taking the 3-set later)
Remove 2-set by looping through discarding any other tile (this creates a number of branches in the simulation)
If removed, return (same as earlier)
Use the number of left-over single tiles to calculate the shanten number
By the way, this is actually very similar to the approach I take when calculating the number myself, and obviously never to yields too high a number.
This works very well for almost all cases. However, I found that sometimes the earlier assumption ("removing already completed 3-sets is NEVER a bad idea") is wrong. Counter-example: 23566M 25667P 159S. The important part is the 25667. By removing a 567 3-set we end up with a left-over 6 tile, leading to 5-shanten. It would be better to use two of the single tiles to form 56x and 67x, leading to 4-shanten overall.
To fix, we simple have to remove the wrong optimization, leading to this code:
Remove completed 3-sets
Remove 2-set by looping through discarding any other tile
Use the number of left-over single tiles to calculate the shanten number
I believe this always accurately finds the smallest shanten number, but I don't know how to prove that. The time taken is in a "reasonable" range (on my machine 10 seconds max, usually 0 seconds).
The final point is calculating the shanten out of the number of left-over single tiles. First of all, it is obvious that the number is in the form 3*n+1 (because we started out with 14 tiles and always subtracted 3 tiles).
If there is 1 tile left, we're shanten already (we're just waiting for the final pair). With 4 tiles left, we have to discard 2 of them to form a 3-set, leaving us with a single tile again. This leads to 2 additional discards. With 7 tiles, we have 2 times 2 discards, adding 4. And so on.
This leads to the simple formula shanten_added = (number_of_singles - 1) * (2/3).
The described algorithm works well and passed all my tests, so I'm assuming it is correct. As stated, I can't prove it though.
Since the algorithm removes the most likely tiles combinations first, it kind of has a built-in optimization. Adding a simple check if (current_depth > best_shanten) then return; it does very well even for high shanten numbers.
My best guess would be an A* inspired approach. You need to find some heuristic which never overestimates the shanten number and use it to search the brute-force tree only in the regions where it is possible to get into a ready state quickly enough.
Correct algorithm sample: syanten.cpp
Recursive cut forms from hand in order: sets, pairs, incomplete forms, - and count it. In all variations. And result is minimal Shanten value of all variants:
Shanten = Min(Shanten, 8 - * 2 - - )
C# sample (rewrited from c++) can be found here (in Russian).
I've done a little bit of thinking and came up with a slightly different formula than mafu's. First of all, consider a hand (a very terrible hand):
1s 4s 6s 1m 5m 8m 9m 9m 7p 8p West East North
By using mafu's algorithm all we can do is cast out a pair (9m,9m). Then we are left with 11 singles. Now if we apply mafu's formula we get (11-1)*2/3 which is not an integer and therefore cannot be a shanten number. This is where I came up with this:
N = ( (S + 1) / 3 ) - 1
N stands for shanten number and S for score sum.
What is score? It's a number of tiles you need to make an incomplete set complete. For example, if you have (4,5) in your hand you need either 3 or 6 to make it a complete 3-set, that is, only one tile. So this incomplete pair gets score 1. Accordingly, (1,1) needs only 1 to become a 3-set. Any single tile obviously needs 2 tiles to become a 3-set and gets score 2. Any complete set of course get score 0. Note that we ignore the possibility of singles becoming pairs. Now if we try to find all of the incomplete sets in the above hand we get:
(4s,6s) (8m,9m) (7p,8p) 1s 1m 5m 9m West East North
Then we count the sum of its scores = 1*3+2*7 = 17.
Now if we apply this number to the formula above we get (17+1)/3 - 1 = 5 which means this hand is 5-shanten. It's somewhat more complicated than Alexey's and I don't have a proof but so far it seems to work for me. Note that such a hand could be parsed in the other way. For example:
(4s,6s) (9m,9m) (7p,8p) 1s 1m 5m 8m West East North
However, it still gets score sum 17 and 5-shanten according to formula. I also can't proof this and this is a little bit more complicated than Alexey's formula but also introduces scores that could be applied(?) to something else.
Take a look here: ShantenNumberCalculator. Calculate shanten really fast. And some related stuff (in japanese, but with code examples) http://cmj3.web.fc2.com
The essence of the algorithm: cut out all pairs, sets and unfinished forms in ALL possible ways, and thereby find the minimum value of the number of shanten.
The maximum value of the shanten for an ordinary hand: 8.
That is, as it were, we have the beginnings for 4 sets and one pair, but only one tile from each (total 13 - 5 = 8).
Accordingly, a pair will reduce the number of shantens by one, two (isolated from the rest) neighboring tiles (preset) will decrease the number of shantens by one,
a complete set (3 identical or 3 consecutive tiles) will reduce the number of shantens by 2, since two suitable tiles came to an isolated tile.
Shanten = 8 - Sets * 2 - Pairs - Presets
Determining whether your hand is already in tenpai sounds like a multi-knapsack problem. Greedy algorithms won't work - as Dialecticus pointed out, you'll need to consider the entire problem space.

Ordering a dictionary to maximize common letters between adjacent words

This is intended to be a more concrete, easily expressable form of my earlier question.
Take a list of words from a dictionary with common letter length.
How to reorder this list tto keep as many letters as possible common between adjacent words?
Example 1:
AGNI, CIVA, DEVA, DEWA, KAMA, RAMA, SIVA, VAYU
reorders to:
AGNI, CIVA, SIVA, DEVA, DEWA, KAMA, RAMA, VAYU
Example 2:
DEVI, KALI, SHRI, VACH
reorders to:
DEVI, SHRI, KALI, VACH
The simplest algorithm seems to be: Pick anything, then search for the shortest distance?
However, DEVI->KALI (1 common) is equivalent to DEVI->SHRI (1 common)
Choosing the first match would result in fewer common pairs in the entire list (4 versus 5).
This seems that it should be simpler than full TSP?
What you're trying to do, is calculate the shortest hamiltonian path in a complete weighted graph, where each word is a vertex, and the weight of each edge is the number of letters that are differenct between those two words.
For your example, the graph would have edges weighted as so:
DEVI KALI SHRI VACH
DEVI X 3 3 4
KALI 3 X 3 3
SHRI 3 3 X 4
VACH 4 3 4 X
Then it's just a simple matter of picking your favorite TSP solving algorithm, and you're good to go.
My pseudo code:
Create a graph of nodes where each node represents a word
Create connections between all the nodes (every node connects to every other node). Each connection has a "value" which is the number of common characters.
Drop connections where the "value" is 0.
Walk the graph by preferring connections with the highest values. If you have two connections with the same value, try both recursively.
Store the output of a walk in a list along with the sum of the distance between the words in this particular result. I'm not 100% sure ATM if you can simply sum the connections you used. See for yourself.
From all outputs, chose the one with the highest value.
This problem is probably NP complete which means that the runtime of the algorithm will become unbearable as the dictionaries grow. Right now, I see only one way to optimize it: Cut the graph into several smaller graphs, run the code on each and then join the lists. The result won't be as perfect as when you try every permutation but the runtime will be much better and the final result might be "good enough".
[EDIT] Since this algorithm doesn't try every possible combination, it's quite possible to miss the perfect result. It's even possible to get caught in a local maximum. Say, you have a pair with a value of 7 but if you chose this pair, all other values drop to 1; if you didn't take this pair, most other values would be 2, giving a much better overall final result.
This algorithm trades perfection for speed. When trying every possible combination would take years, even with the fastest computer in the world, you must find some way to bound the runtime.
If the dictionaries are small, you can simply create every permutation and then select the best result. If they grow beyond a certain bound, you're doomed.
Another solution is to mix the two. Use the greedy algorithm to find "islands" which are probably pretty good and then use the "complete search" to sort the small islands.
This can be done with a recursive approach. Pseudo-code:
Start with one of the words, call it w
FindNext(w, l) // l = list of words without w
Get a list l of the words near to w
If only one word in list
Return that word
Else
For every word w' in l do FindNext(w', l') //l' = l without w'
You can add some score to count common pairs and to prefer "better" lists.
You may want to take a look at BK-Trees, which make finding words with a given distance to each other efficient. Not a total solution, but possibly a component of one.
This problem has a name: n-ary Gray code. Since you're using English letters, n = 26. The Wikipedia article on Gray code describes the problem and includes some sample code.

Resources