How to solve the "Mastermind" guessing game?

Question

How would you create an algorithm to solve the following puzzle, "Mastermind"?

Your opponent has chosen four different colours from a set of six (yellow, blue, green, red, orange, purple). You must guess which they have chosen, and in what order. After each guess, your opponent tells you how many (but not which) of the colours you guessed were the right colour in the right place ["blacks"] and how many (but not which) were the right colour but in the wrong place ["whites"]. The game ends when you guess correctly (4 blacks, 0 whites).

For example, if your opponent has chosen (blue, green, orange, red), and you guess (yellow, blue, green, red), you will get one "black" (for the red), and two whites (for the blue and green). You would get the same score for guessing (blue, orange, red, purple).

I'm interested in what algorithm you would choose, and (optionally) how you translate that into code (preferably Python). I'm interested in coded solutions that are:

1. Clear (easily understood)
2. Concise
3. Efficient (fast in making a guess)
4. Effective (least number of guesses to solve the puzzle)
5. Flexible (can easily answer questions about the algorithm, e.g. what is its worst case?)
6. General (can be easily adapted to other types of puzzle than Mastermind)

I'm happy with an algorithm that's very effective but not very efficient (provided it's not just poorly implemented!); however, a very efficient and effective algorithm implemented inflexibly and impenetrably is not of use.

I have my own (detailed) solution in Python which I have posted, but this is by no means the only or best approach, so please post more! I'm not expecting an essay ;)

Answer 1

Key tools: entropy, greediness, branch-and-bound; Python, generators, itertools, decorate-undecorate pattern

In answering this question, I wanted to build up a language of useful functions to explore the problem. I will go through these functions, describing them and their intent. Originally, these had extensive docs, with small embedded unit tests tested using doctest; I can't praise this methodology highly enough as a brilliant way to implement test-driven-development. However, it does not translate well to StackOverflow, so I will not present it this way.

Firstly, I will be needing several standard modules and future imports (I work with Python 2.6).

from __future__ import division # No need to cast to float when dividing import collections, itertools, math 

I will need a scoring function. Originally, this returned a tuple (blacks, whites), but I found output a little clearer if I used a namedtuple:

Pegs = collections.namedtuple('Pegs', 'black white') def mastermindScore(g1,g2):   matching = len(set(g1) & set(g2))   blacks = sum(1 for v1, v2 in itertools.izip(g1,g2) if v1 == v2)   return Pegs(blacks, matching-blacks) 

To make my solution general, I pass in anything specific to the Mastermind problem as keyword arguments. I have therefore made a function that creates these arguments once, and use the **kwargs syntax to pass it around. This also allows me to easily add new attributes if I need them later. Note that I allow guesses to contain repeats, but constrain the opponent to pick distinct colours; to change this, I only need change G below. (If I wanted to allow repeats in the opponent's secret, I would need to change the scoring function as well.)

def mastermind(colours, holes):   return dict(     G           = set(itertools.product(colours,repeat=holes)),     V           = set(itertools.permutations(colours, holes)),     score       = mastermindScore,     endstates   = (Pegs(holes, 0),))  def mediumGame():     return mastermind(("Yellow", "Blue", "Green", "Red", "Orange", "Purple"), 4) 

Sometimes I will need to partition a set based on the result of applying a function to each element in the set. For instance, the numbers 1..10 can be partitioned into even and odd numbers by the function n % 2 (odds give 1, evens give 0). The following function returns such a partition, implemented as a map from the result of the function call to the set of elements that gave that result (e.g. { 0: evens, 1: odds }).

def partition(S, func, *args, **kwargs):   partition = collections.defaultdict(set)   for v in S: partition[func(v, *args, **kwargs)].add(v)   return partition 

I decided to explore a solver that uses a greedy entropic approach. At each step, it calculates the information that could be obtained from each possible guess, and selects the most informative guess. As the numbers of possibilities grow, this will scale badly (quadratically), but let's give it a try! First, I need a method to calculate the entropy (information) of a set of probabilities. This is just -∑p log p. For convenience, however, I will allow input that are not normalized, i.e. do not add up to 1:

def entropy(P):   total = sum(P)   return -sum(p*math.log(p, 2) for p in (v/total for v in P if v)) 

So how am I going to use this function? Well, for a given set of possibilities, V, and a given guess, g, the information we get from that guess can only come from the scoring function: more specifically, how that scoring function partitions our set of possibilities. We want to make a guess that distinguishes best among the remaining possibilites — divides them into the largest number of small sets — because that means we are much closer to the answer. This is exactly what the entropy function above is putting a number to: a large number of small sets will score higher than a small number of large sets. All we need to do is plumb it in.

def decisionEntropy(V, g, score):   return entropy(collections.Counter(score(gi, g) for gi in V).values()) 

Of course, at any given step what we will actually have is a set of remaining possibilities, V, and a set of possible guesses we could make, G, and we will need to pick the guess which maximizes the entropy. Additionally, if several guesses have the same entropy, prefer to pick one which could also be a valid solution; this guarantees the approach will terminate. I use the standard python decorate-undecorate pattern together with the built-in max method to do this:

def bestDecision(V, G, score):   return max((decisionEntropy(V, g, score), g in V, g) for g in G)[2] 

Now all I need to do is repeatedly call this function until the right result is guessed. I went through a number of implementations of this algorithm until I found one that seemed right. Several of my functions will want to approach this in different ways: some enumerate all possible sequences of decisions (one per guess the opponent may have made), while others are only interested in a single path through the tree (if the opponent has already chosen a secret, and we are just trying to reach the solution). My solution is a "lazy tree", where each part of the tree is a generator that can be evaluated or not, allowing the user to avoid costly calculations they won't need. I also ended up using two more namedtuples, again for clarity of code.

Node = collections.namedtuple('Node', 'decision branches') Branch = collections.namedtuple('Branch', 'result subtree') def lazySolutionTree(G, V, score, endstates, **kwargs):   decision = bestDecision(V, G, score)   branches = (Branch(result, None if result in endstates else                    lazySolutionTree(G, pV, score=score, endstates=endstates))               for (result, pV) in partition(V, score, decision).iteritems())   yield Node(decision, branches) # Lazy evaluation 

The following function evaluates a single path through this tree, based on a supplied scoring function:

def solver(scorer, **kwargs):   lazyTree = lazySolutionTree(**kwargs)   steps = []   while lazyTree is not None:     t = lazyTree.next() # Evaluate node     result = scorer(t.decision)     steps.append((t.decision, result))     subtrees = [b.subtree for b in t.branches if b.result == result]     if len(subtrees) == 0:       raise Exception("No solution possible for given scores")     lazyTree = subtrees[0]   assert(result in endstates)   return steps 

This can now be used to build an interactive game of Mastermind where the user scores the computer's guesses. Playing around with this reveals some interesting things. For example, the most informative first guess is of the form (yellow, yellow, blue, green), not (yellow, blue, green, red). Extra information is gained by using exactly half the available colours. This also holds for 6-colour 3-hole Mastermind — (yellow, blue, green) — and 8-colour 5-hole Mastermind — (yellow, yellow, blue, green, red).

But there are still many questions that are not easily answered with an interactive solver. For instance, what is the most number of steps needed by the greedy entropic approach? And how many inputs take this many steps? To make answering these questions easier, I first produce a simple function that turns the lazy tree of above into a set of paths through this tree, i.e. for each possible secret, a list of guesses and scores.

def allSolutions(**kwargs):   def solutions(lazyTree):     return ((((t.decision, b.result),) + solution              for t in lazyTree for b in t.branches              for solution in solutions(b.subtree))             if lazyTree else ((),))   return solutions(lazySolutionTree(**kwargs)) 

Finding the worst case is a simple matter of finding the longest solution:

def worstCaseSolution(**kwargs):   return max((len(s), s) for s in allSolutions(**kwargs)) [1] 

It turns out that this solver will always complete in 5 steps or fewer. Five steps! I know that when I played Mastermind as a child, I often took longer than this. However, since creating this solver and playing around with it, I have greatly improved my technique, and 5 steps is indeed an achievable goal even when you don't have time to calculate the entropically ideal guess at each step ;)

How likely is it that the solver will take 5 steps? Will it ever finish in 1, or 2, steps? To find that out, I created another simple little function that calculates the solution length distribution:

def solutionLengthDistribution(**kwargs):   return collections.Counter(len(s) for s in allSolutions(**kwargs)) 

For the greedy entropic approach, with repeats allowed: 7 cases take 2 steps; 55 cases take 3 steps; 229 cases take 4 steps; and 69 cases take the maximum of 5 steps.

Of course, there's no guarantee that the greedy entropic approach minimizes the worst-case number of steps. The final part of my general-purpose language is an algorithm that decides whether or not there are any solutions for a given worst-case bound. This will tell us whether greedy entropic is ideal or not. To do this, I adopt a branch-and-bound strategy:

def solutionExists(maxsteps, G, V, score, **kwargs):   if len(V) == 1: return True   partitions = [partition(V, score, g).values() for g in G]   maxSize = max(len(P) for P in partitions) ** (maxsteps - 2)   partitions = (P for P in partitions if max(len(s) for s in P) <= maxSize)   return any(all(solutionExists(maxsteps-1,G,s,score) for l,s in                  sorted((-len(s), s) for s in P)) for i,P in              sorted((-entropy(len(s) for s in P), P) for P in partitions)) 

This is definitely a complex function, so a bit more explanation is in order. The first step is to partition the remaining solutions based on their score after a guess, as before, but this time we don't know what guess we're going to make, so we store all partitions. Now we could just recurse into every one of these, effectively enumerating the entire universe of possible decision trees, but this would take a horrifically long time. Instead I observe that, if at this point there is no partition that divides the remaining solutions into more than n sets, then there can be no such partition at any future step either. If we have k steps left, that means we can distinguish between at most n^k-1 solutions before we run out of guesses (on the last step, we must always guess correctly). Thus we can discard any partitions that contain a score mapped to more than this many solutions. This is the next two lines of code.

The final line of code does the recursion, using Python's any and all functions for clarity, and trying the highest-entropy decisions first to hopefully minimize runtime in the positive case. It also recurses into the largest part of the partition first, as this is the most likely to fail quickly if the decision was wrong. Once again, I use the standard decorate-undecorate pattern, this time to wrap Python's sorted function.

def lowerBoundOnWorstCaseSolution(**kwargs):   for steps in itertools.count(1):     if solutionExists(maxsteps=steps, **kwargs):       return steps 

By calling solutionExists repeatedly with an increasing number of steps, we get a strict lower bound on the number of steps needed in the worst case for a Mastermind solution: 5 steps. The greedy entropic approach is indeed optimal.

Out of curiosity, I invented another guessing game, which I nicknamed "twoD". In this, you try to guess a pair of numbers; at each step, you get told if your answer is correct, if the numbers you guessed are no less than the corresponding ones in the secret, and if the numbers are no greater.

Comparison = collections.namedtuple('Comparison', 'less greater equal') def twoDScorer(x, y):   return Comparison(all(r[0] <= r[1] for r in zip(x, y)),                     all(r[0] >= r[1] for r in zip(x, y)),                     x == y) def twoD():   G = set(itertools.product(xrange(5), repeat=2))   return dict(G = G, V = G, score = twoDScorer,               endstates = set(Comparison(True, True, True))) 

For this game, the greedy entropic approach has a worst case of five steps, but there is a better solution possible with a worst case of four steps, confirming my intuition that myopic greediness is only coincidentally ideal for Mastermind. More importantly, this has shown how flexible my language is: all the same methods work for this new guessing game as did for Mastermind, letting me explore other games with a minimum of extra coding.

What about performance? Obviously, being implemented in Python, this code is not going to be blazingly fast. I've also dropped some possible optimizations in favour of clear code.

One cheap optimization is to observe that, on the first move, most guesses are basically identical: (yellow, blue, green, red) is really no different from (blue, red, green, yellow), or (orange, yellow, red, purple). This greatly reduces the number of guesses we need consider on the first step — otherwise the most costly decision in the game.

However, because of the large runtime growth rate of this problem, I was not able to solve the 8-colour, 5-hole Mastermind problem, even with this optimization. Instead, I ported the algorithms to C++, keeping the general structure the same and employing bitwise operations to boost performance in the critical inner loops, for a speedup of many orders of magnitude. I leave this as an exercise to the reader :)

Addendum, 2018: It turns out the greedy entropic approach is not optimal for the 8-colour, 4-hole Mastermind problem either, with a worst-case length of 7 steps when an algorithm exists that takes at most 6!

Answer 2

I once wrote a "Jotto" solver which is essentially "Master Mind" with words. (We each pick a word and we take turns guessing at each other's word, scoring "right on" (exact) matches and "elsewhere" (correct letter/color, but wrong placement).

The key to solving such a problem is the realization that the scoring function is symmetric.

In other words if score(myguess) == (1,2) then I can use the same score() function to compare my previous guess with any other possibility and eliminate any that don't give exactly the same score.

Let me give an example: The hidden word (target) is "score" ... the current guess is "fools" --- the score is 1,1 (one letter, 'o', is "right on"; another letter, 's', is "elsewhere"). I can eliminate the word "guess" because the `score("guess") (against "fools") returns (1,0) (the final 's' matches, but nothing else does). So the word "guess" is not consistent with "fools" and a score against some unknown word that gave returned a score of (1,1).

So I now can walk through every five letter word (or combination of five colors/letters/digits etc) and eliminate anything that doesn't score 1,1 against "fools." Do that at each iteration and you'll very rapidly converge on the target. (For five letter words I was able to get within 6 tries every time ... and usually only 3 or 4). Of course there's only 6000 or so "words" and you're eliminating close to 95% for each guess.

Note: for the following discussion I'm talking about five letter "combination" rather than four elements of six colors. The same algorithms apply; however, the problem is orders of magnitude smaller for the old "Master Mind" game ... there are only 1296 combinations (6**4) of colored pegs in the classic "Master Mind" program, assuming duplicates are allowed. The line of reasoning that leads to the convergence involves some combinatorics: there are 20 non-winning possible scores for a five element target (n = [(a,b) for a in range(5) for b in range(6) if a+b <= 5] to see all of them if you're curious. We would, therefore, expect that any random valid selection would have a roughly 5% chance of matching our score ... the other 95% won't and therefore will be eliminated for each scored guess. This doesn't account for possible clustering in word patterns but the real world behavior is close enough for words and definitely even closer for "Master Mind" rules. However, with only 6 colors in 4 slots we only have 14 possible non-winning scores so our convergence isn't quite as fast).

For Jotto the two minor challenges are: generating a good world list (awk -f 'length($0)==5' /usr/share/dict/words or similar on a UNIX system) and what to do if the user has picked a word that not in our dictionary (generate every letter combination, 'aaaaa' through 'zzzzz' --- which is 26 ** 5 ... or ~1.1 million). A trivial combination generator in Python takes about 1 minute to generate all those strings ... an optimized one should to far better. (I can also add a requirement that every "word" have at least one vowel ... but this constraint doesn't help much --- 5 vowels * 5 possible locations for that and then multiplied by 26 ** 4 other combinations).

For Master Mind you use the same combination generator ... but with only 4 or 5 "letters" (colors). Every 6-color combination (15,625 of them) can be generated in under a second (using the same combination generator as I used above).

If I was writing this "Jotto" program today, in Python for example, I would "cheat" by having a thread generating all the letter combos in the background while I was still eliminated words from the dictionary (while my opponent was scoring me, guessing, etc). As I generated them I'd also eliminate against all guesses thus far. Thus I would, after I'd eliminated all known words, have a relatively small list of possibilities and against a human player I've "hidden" most of my computation lag by doing it in parallel to their input. (And, if I wrote a web server version of such a program I'd have my web engine talk to a local daemon to ask for sequences consistent with a set of scores. The daemon would keep a locally generated list of all letter combinations and would use a select.select() model to feed possibilities back to each of the running instances of the game --- each would feed my daemon word/score pairs which my daemon would apply as a filter on the possibilities it feeds back to that client).

(By comparison I wrote my version of "Jotto" about 20 years ago on an XT using Borland TurboPascal ... and it could do each elimination iteration --- starting with its compiled in list of a few thousand words --- in well under a second. I build its word list by writing a simple letter combination generator (see below) ... saving the results to a moderately large file, then running my word processor's spell check on that with a macro to delete everything that was "mis-spelled" --- then I used another macro to wrap all the remaining lines in the correct punctuation to make them valid static assignments to my array, which was a #include file to my program. All that let me build a standalone game program that "knew" just about every valid English 5 letter word; the program was a .COM --- less than 50KB if I recall correctly).

For other reasons I've recently written a simple arbitrary combination generator in Python. It's about 35 lines of code and I've posted that to my "trite snippets" wiki on bitbucket.org ... it's not a "generator" in the Python sense ... but a class you can instantiate to an infinite sequence of "numeric" or "symbolic" combination of elements (essentially counting in any positive integer base).

You can find it at: Trite Snippets: Arbitrary Sequence Combination Generator

For the exact match part of our score() function you can just use this:

def score(this, that):     '''Simple "Master Mind" scoring function'''     exact = len([x for x,y in zip(this, that) if x==y])     ### Calculating "other" (white pegs) goes here:     ### ...     ###     return (exact,other) 

I think this exemplifies some of the beauty of Python: zip() up the two sequences, return any that match, and take the length of the results).

Finding the matches in "other" locations is deceptively more complicated. If no repeats were allowed then you could simply use sets to find the intersections.

[In my earlier edit of this message, when I realized how I could use zip() for exact matches, I erroneously thought we could get away with other = len([x for x,y in zip(sorted(x), sorted(y)) if x==y]) - exact ... but it was late and I was tired. As I slept on it I realized that the method was flawed. Bad, Jim! Don't post without adequate testing!* (Tested several cases that happened to work)].

In the past the approach I used was to sort both lists, compare the heads of each: if the heads are equal, increment the count and pop new items from both lists. otherwise pop a new value into the lesser of the two heads and try again. Break as soon as either list is empty.

This does work; but it's fairly verbose. The best I can come up with using that approach is just over a dozen lines of code:

other = 0 x = sorted(this)   ## Implicitly converts to a list! y = sorted(that) while len(x) and len(y):     if x[0] == y[0]:         other += 1         x.pop(0)         y.pop(0)     elif x[0] < y[0]:         x.pop(0)     else:         y.pop(0) other -= exact 

Using a dictionary I can trim that down to about nine:

other = 0 counters = dict() for i in this:     counters[i] = counters.get(i,0) + 1 for i in that:     if counters.get(i,0) > 0:         other += 1         counters[i] -= 1 other -= exact 

(Using the new "collections.Counter" class (Python3 and slated for Python 2.7?) I could presumably reduce this a little more; three lines here are initializing the counters collection).

It's important to decrement the "counter" when we find a match and it's vital to test for counter greater than zero in our test. If a given letter/symbol appears in "this" once and "that" twice then it must only be counted as a match once.

The first approach is definitely a bit trickier to write (one must be careful to avoid boundaries). Also in a couple of quick benchmarks (testing a million randomly generated pairs of letter patterns) the first approach takes about 70% longer as the one using dictionaries. (Generating the million pairs of strings using random.shuffle() took over twice as long as the slower of the scoring functions, on the other hand).

A formal analysis of the performance of these two functions would be complicated. The first method has two sorts, so that would be 2 * O(nlog(n)) ... and it iterates through at least one of the two strings and possibly has to iterate all the way to the end of the other string (best case O(n), worst case O(2n)) -- force I'm mis-using big-O notation here, but this is just a rough estimate. The second case depends entirely on the perfomance characteristics of the dictionary. If we were using b-trees then the performance would be roughly O(nlog(n) for creation and finding each element from the other string therein would be another O(n*log(n)) operation. However, Python dictionaries are very efficient and these operations should be close to constant time (very few hash collisions). Thus we'd expect a performance of roughly O(2n) ... which of course simplifies to O(n). That roughly matches my benchmark results.

Glancing over the Wikipedia article on "Master Mind" I see that Donald Knuth used an approach which starts similarly to mine (and 10 years earlier) but he added one significant optimization. After gathering every remaining possibility he selects whichever one would eliminate the largest number of possibilities on the next round. I considered such an enhancement to my own program and rejected the idea for practical reasons. In his case he was searching for an optimal (mathematical) solution. In my case I was concerned about playability (on an XT, preferably using less than 64KB of RAM, though I could switch to .EXE format and use up to 640KB). I wanted to keep the response time down in the realm of one or two seconds (which was easy with my approach but which would be much more difficult with the further speculative scoring). (Remember I was working in Pascal, under MS-DOS ... no threads, though I did implement support for crude asynchronous polling of the UI which turned out to be unnecessary)

If I were writing such a thing today I'd add a thread to do the better selection, too. This would allow me to give the best guess I'd found within a certain time constraint, to guarantee that my player didn't have to wait too long for my guess. Naturally my selection/elimination would be running while waiting for my opponent's guesses.