I'm self-learning with the edX course CS 188.1x Artificial Intelligence. Since the course concluded a year ago, it is in "archive mode" and there are no teaching staff to help with the questions. This also means I get no credit for finishing the course, so hopefully asking for help with a "homework" question here is okay.
In the first homework assignment the following question is asked:
Question 9: Hive Minds Lost at Night It is night and you control a single insect. You know the maze, but you do not know what square the
insect will start in. You must pose a search problem whose solution is
an all-purpose sequence of actions such that, after executing those
actions, the insect will be on the exit square, regardless of initial
position. The insect executes the actions mindlessly and does not know
whether its moves succeed: if it uses an action which would move it in
a blocked direction, it will stay where it is. For example, in the
maze below, moving right twice guarantees that the insect will be at
the exit regardless of its starting position.
It then asks the size of the state space. The answer is given as 2^MN, where M and N are the horizontal and vertical dimensions of the maze. Why is the answer the power set of MN? In my mind, the bug can only be in one square at the beginning, and we only have one bug, so I know the number of start states is MN. But the number of start states != state space, and that is where I am confused.
FYI - the cost per move is 1, and the bug can only move 1 square left, right, up, or down at a time. The goal is to get to the X (goal square).
Okay - I think I got it.
Set of all subsets (power set*) is exactly the right way to think about this. The state space is the set of all states.
1) Definition of state:
"A state contains all of the information necessary to predict the effects of an action and to
determine if it is a goal state." (http://artint.info/html/ArtInt_48.html)
The actions in this scenario are simple: left, right, up, down. They are the possible movements a bug could make.
2) Definition of solution:
Solutions are sequences of actions that lead to the goal test being
passed.
If we only permitted MN states, one for each possible starting position the bug was in, then we would have a state space that gave solutions that were valid only for discrete starting positions. But, the solution must be valid regardless of the initial state of the bug. This means the solution must work for scenarios in which the bug could occupy any of the MN available squares.
In other words, the solutions must be valid for each and every subset (combination) of possible starting spaces, which yields the power set of MN, which is 2^MN.
Why? Because solutions that are valid for a given start state may not be valid for all other start states. And the problem requires us to find solutions that are valid for all start states. This is also why the state space is much larger than MN even though in reality our bug only occupies 1 of the MN positions upon initialization. Just because a solution (sequence of moves) works when the bug starts at (1, 1) doesn't mean that solution (sequence of moves) will also work for the bug starting at (2, 1).
Bonus Question: Why isn't the state space just 1, the full set where
each of the MN squares 'has' a bug (and bugs are permitted to move
on top of each other)?
I was tempted to say that just because a sequence of moves gives a goal state when the bug can start at all of the MN possible positions, that doesn't mean that same sequence of moves gives a goal state when the bug starts at (3, 2) or at MN - 1 or MN - 2 etc. possible positions. But by definition it must (a solution over all starting points must be a solution over every finite subset of starting points).
So I think the reason you evaluate starting states other than "all boxes have a bug" is because the solution generated by evaluating only that state may not be optimal. And in fact this interpretation is borne out by what the homework gives as admissible heuristics for this problem:
The maximum of Manhattan distances to the goal from each possible
location the insect could be in.
OR
The minimum of Manhattan distances to the goal from each possible
location the insect could be in.
The case where we just have one starting state with bugs on all the boxes (with the magic ability to be on top of each other) is the relaxed problem we use to define our heuristic. Again by definition of admissibility, since the heuristic must not overestimate the true (arc) cost of an action, and since arc cost is given by Manhattan distances, both the heuristics above are admissible. (The maximum case is admissible because each possible location for the bug is, in fact, possible - thus the max cost is possible).
*If you don't know what power set means, all you need to know is that the power set is the set of all subsets of a given set. It is given by 2^(size of the set).
In other words, if I have a set of three balls {red, blue, green} and I want to know how many different subsets I have, I can calculate it as follows. A subset either has the element in it (1), or it doesn't (0). So {0, 0, 1} would be the subset of only the green ball, {1, 1, 1} would be the subset of all the balls (yes, technically this is a subset) and {0, 0, 0} would be the subset of none of the balls (again, technically a subset). So we see that the number of all subsets is 2^3 = 8. Or in our problem, 2^MN.
Related
I am trying to understand the example from the book https://cses.fi/book/book.pdf at page 239 .
The example is described as follows:
What I don't get is just what exactly, say, number 3 next to lower right corner means, we can move 4 steps up and 3 steps left from it, how is it 3? Same for 4 just above it, it doesn't correspond to any set of moves I can think of. The book in general makes a lot of leaps of logic they think are obvious but usually I can infer what they mean after some time, here I am just lost.
The rule for computing these numbers is recursive.
You consider all the values you can reach, and then pick the smallest (non-negative) integer that is not reachable.
For example, the value in the top-left corner is 0 because no moves are possible.
For example, the value next to lower right is 3 because the reachable values are 0,4,1,0,2,1,4 so 3 is the smallest integer not in this list.
This explains how to compute the numbers, but to understand them it is probably better to start with understanding the game of Nim. In the game of Nim, the Sprague Grundy number for a pile is simply equal to the size of a pile.
the unknown one blue marked
it's 3/5? 1/3? 2/5? or max value of above, or maybe another(I think this...)?
how to caculate? it confuses me very much...
It's actually 1/2.
Note that there must be at least two mines in the three cells below 3 (because the other two cells are adjacent to a 1 and as such can't have more than 1 mine).
This means that there must be at least one mine between the cell below the 3 and cell to the bottom right of 3. Since both of these are also adjacent to a 1, only one of them can have mine at most. Thus, exactly one of these cells is a mine. The 2 on the right becomes irrelevant at this point.
With that out of the way, is there a general algorithm which can generate results like these?
I can't thing of any polynomial time solutions but it might be possible to simple try all the alternatives while backtracking when a constraint fails.
Consider the following linked list:
1->2->3->4->5->6->7->8->9->4->...->9->4.....
The above list has a loop as follows:
[4->5->6->7->8->9->4]
Drawing the linked list on a whiteboard, I tried manually solving it for different pointer steps, to see how the pointers move around -
(slow_pointer_increment, fast_pointer_increment)
So, the pointers for different cases are as follows:
(1,2), (2,3), (1,3)
The first two pairs of increments - (1,2) and (2,3) worked fine, but when I use the pair (1,3), the algorithm does not seem to work on this pair. Is there a rule as to by how much we need to increment the steps for this algorithm to hold true?
Although I searched for various increment steps for the slower and the faster pointer, I haven't so far found a single relevant answer as to why it is not working for the increment (1,3) on this list.
The algorithm can easily be shown to be guaranteed to find a cycle starting from any position if the difference between the pointer increments and the cycle length are coprimes (i.e. their greatest common divisor must be 1).
For the general case, this means the difference between the increments must be 1 (because that's the only positive integer that's coprime to all other positive integers).
For any given pointer increments, if the values aren't coprimes, it may still be guaranteed to find a cycle, but one would need to come up with a different way to prove that it will find a cycle.
For the example in the question, with pointer increments of (1,3), the difference is 3-1=2, and the cycle length is 6. 2 and 6 are not coprimes, thus it's not known whether the algorithm is guaranteed to find the cycle in general. It does seem like this might actually be guaranteed to find the cycle (including for the example in the question), even though it doesn't reach every position (which applies with coprime, as explained below), but I don't have a proof for this at the moment.
The key to understanding this is that, at least for the purposes of checking whether the pointers ever meet, the slow and fast pointers' positions within the cycle only matters relative to each other. That is, these two can be considered equivalent: (the difference is 1 for both)
slow fast slow fast
↓ ↓ ↓ ↓
0→1→2→3→4→5→0 0→1→2→3→4→5→0
So we can think of this in terms of the position of slow remaining constant and fast moving at an increment of fastIncrement-slowIncrement, at which point the problem becomes:
Starting at any position, can we reach a specific position moving at some speed (mod cycle length)?
Or, more generally:
Can we reach every position moving at some speed (mod cycle length)?
Which will only be true if the speed and cycle length are coprimes.
For example, look at a speed of 4 and a cycle of length 6 - starting at 0, we visit:
0, 4, 8%6=2, 6%6=0, 4, 2, 0, ... - GCD(4,6) = 2, and we can only visit every second element.
To see this in action, consider increments of (1,5) (difference = 4) for the example given above and see that the pointers will never meet.
I should note that, to my knowledge at least, the (1,2) increment is considered a fundamental part of the algorithm.
Using different increments (as per the above constraints) might work, but it would be a move away from the "official" algorithm and would involve more work (since a pointer to a linked-list must be incremented iteratively, you can't increment it by more than 1 in a single step) without any clear advantage for the general case.
Bernhard Barker explanation is spot on.
I am simply adding on to it.
Why should the difference of speeds between the pointers and the cycle length be
coprime numbers?
Take a scenario where the difference of speeds between pointers(say v) and cycle length(say L) are not coprime.
So there exists a GCD(v,L) greater than 1 (say G).
Therefore, we have
v=difference of speeds between pointers
L=Length of the cycle(i.e. the number of nodes in the cycle)
G=GCD(v,L)
Since we are considering only relative positions, essentially the slow is stationary and the fast is moving at a relative speed v.
Let fast be at some node in the cycle.
Since G is a divisor of L we can divide the cycle into G/L parts. Start dividing from where fast is located.
Now, v is a multiple of G (say v=nG).
Every time the fast pointer moves it will jump across n parts. So in each part the pointer arrives on a single node(basically the last node of a part). Each and every time the fast pointer will land on the ending node of every part. Refer the image below
Example image
As mentioned above by Bernhard, the question we need to answer is
Can we reach every position moving at some speed?
The answer is no if we have a GCD existing. As we see the fast pointer will only cover the last nodes in every part.
I need some advice. I'm developing a game similar to Flow Free wherein the gameboard is composed of a grid and colored dots, and the user has to connect the same colored dots together without overlapping other lines, and using up ALL the free spaces in the board.
My question is about level-creation. I wish to make the levels generated randomly (and should at least be able to solve itself so that it can give players hints) and I am in a stump as to what algorithm to use. Any suggestions?
Note: image shows the objective of Flow Free, and it is the same objective of what I am developing.
Thanks for your help. :)
Consider solving your problem with a pair of simpler, more manageable algorithms: one algorithm that reliably creates simple, pre-solved boards and another that rearranges flows to make simple boards more complex.
The first part, building a simple pre-solved board, is trivial (if you want it to be) if you're using n flows on an nxn grid:
For each flow...
Place the head dot at the top of the first open column.
Place the tail dot at the bottom of that column.
Alternatively, you could provide your own hand-made starter boards to pass to the second part. The only goal of this stage is to get a valid board built, even if it's just trivial or predetermined, so it's worth keeping it simple.
The second part, rearranging the flows, involves looping over each flow, seeing which one can work with its neighboring flow to grow and shrink:
For some number of iterations...
Choose a random flow f.
If f is at the minimum length (say 3 squares long), skip to the next iteration because we can't shrink f right now.
If the head dot of f is next to a dot from another flow g (if more than one g to choose from, pick one at random)...
Move f's head dot one square along its flow (i.e., walk it one square towards the tail). f is now one square shorter and there's an empty square. (The puzzle is now unsolved.)
Move the neighboring dot from g into the empty square vacated by f. Now there's an empty square where g's dot moved from.
Fill in that empty spot with flow from g. Now g is one square longer than it was at the beginning of this iteration. (The puzzle is back to being solved as well.)
Repeat the previous step for f's tail dot.
The approach as it stands is limited (dots will always be neighbors) but it's easy to expand upon:
Add a step to loop through the body of flow f, looking for trickier ways to swap space with other flows...
Add a step that prevents a dot from moving to an old location...
Add any other ideas that you come up with.
The overall solution here is probably less than the ideal one that you're aiming for, but now you have two simple algorithms that you can flesh out further to serve the role of one large, all-encompassing algorithm. In the end, I think this approach is manageable, not cryptic, and easy to tweek, and, if nothing else, a good place to start.
Update: I coded a proof-of-concept based on the steps above. Starting with the first 5x5 grid below, the process produced the subsequent 5 different boards. Some are interesting, some are not, but they're always valid with one known solution.
Starting Point
5 Random Results (sorry for the misaligned screenshots)
And a random 8x8 for good measure. The starting point was the same simple columns approach as above.
Updated answer: I implemented a new generator using the idea of "dual puzzles". This allows much sparser and higher quality puzzles than any previous method I know of. The code is on github. I'll try to write more details about how it works, but here is an example puzzle:
Old answer:
I have implemented the following algorithm in my numberlink solver and generator. In enforces the rule, that a path can never touch itself, which is normal in most 'hardcore' numberlink apps and puzzles
First the board is tiled with 2x1 dominos in a simple, deterministic way.
If this is not possible (on an odd area paper), the bottom right corner is
left as a singleton.
Then the dominos are randomly shuffled by rotating random pairs of neighbours.
This is is not done in the case of width or height equal to 1.
Now, in the case of an odd area paper, the bottom right corner is attached to
one of its neighbour dominos. This will always be possible.
Finally, we can start finding random paths through the dominos, combining them
as we pass through. Special care is taken not to connect 'neighbour flows'
which would create puzzles that 'double back on themselves'.
Before the puzzle is printed we 'compact' the range of colours used, as much as possible.
The puzzle is printed by replacing all positions that aren't flow-heads with a .
My numberlink format uses ascii characters instead of numbers. Here is an example:
$ bin/numberlink --generate=35x20
Warning: Including non-standard characters in puzzle
35 20
....bcd.......efg...i......i......j
.kka........l....hm.n....n.o.......
.b...q..q...l..r.....h.....t..uvvu.
....w.....d.e..xx....m.yy..t.......
..z.w.A....A....r.s....BB.....p....
.D.........E.F..F.G...H.........IC.
.z.D...JKL.......g....G..N.j.......
P...a....L.QQ.RR...N....s.....S.T..
U........K......V...............T..
WW...X.......Z0..M.................
1....X...23..Z0..........M....44...
5.......Y..Y....6.........C.......p
5...P...2..3..6..VH.......O.S..99.I
........E.!!......o...."....O..$$.%
.U..&&..J.\\.(.)......8...*.......+
..1.......,..-...(/:.."...;;.%+....
..c<<.==........)./..8>>.*.?......#
.[..[....]........:..........?..^..
..._.._.f...,......-.`..`.7.^......
{{......].....|....|....7.......#..
And here I run it through my solver (same seed):
$ bin/numberlink --generate=35x20 | bin/numberlink --tubes
Found a solution!
┌──┐bcd───┐┌──efg┌─┐i──────i┌─────j
│kka│└───┐││l┌─┘│hm│n────n┌o│┌────┐
│b──┘q──q│││l│┌r└┐│└─h┌──┐│t││uvvu│
└──┐w┌───┘d└e││xx│└──m│yy││t││└──┘│
┌─z│w│A────A┌┘└─r│s───┘BB││┌┘└p┌─┐│
│D┐└┐│┌────E│F──F│G──┐H┐┌┘││┌──┘IC│
└z└D│││JKL┌─┘┌──┐g┌─┐└G││N│j│┌─┐└┐│
P──┐a││││L│QQ│RR└┐│N└──┘s││┌┘│S│T││
U─┐│┌┘││└K└─┐└─┐V││└─────┘││┌┘││T││
WW│││X││┌──┐│Z0││M│┌──────┘││┌┘└┐││
1┐│││X│││23││Z0│└┐││┌────M┌┘││44│││
5│││└┐││Y││Y│┌─┘6││││┌───┐C┌┘│┌─┘│p
5││└P│││2┘└3││6─┘VH│││┌─┐│O┘S┘│99└I
┌┘│┌─┘││E┐!!│└───┐o┘│││"│└─┐O─┘$$┌%
│U┘│&&│└J│\\│(┐)┐└──┘│8││┌*└┐┌───┘+
└─1└─┐└──┘,┐│-└┐│(/:┌┘"┘││;;│%+───┘
┌─c<<│==┌─┐││└┐│)│/││8>>│*┌?│┌───┐#
│[──[└─┐│]││└┐│└─┘:┘│└──┘┌┘┌┘?┌─^││
└─┐_──_│f││└,│└────-│`──`│7┘^─┘┌─┘│
{{└────┘]┘└──┘|────|└───7└─────┘#─┘
I've tested replacing step (4) with a function that iteratively, randomly merges two neighboring paths. However it game much denser puzzles, and I already think the above is nearly too dense to be difficult.
Here is a list of problems I've generated of different size: https://github.com/thomasahle/numberlink/blob/master/puzzles/inputs3
The most straightforward way to create such a level is to find a way to solve it. This way, you can basically generate any random starting configuration and determine if it is a valid level by trying to have it solved. This will generate the most diverse levels.
And even if you find a way to generate the levels some other way, you'll still want to apply this solving algorithm to prove that the generated level is any good ;)
Brute-force enumerating
If the board has a size of NxN cells, and there are also N colours available, brute-force enumerating all possible configurations (regardless of wether they form actual paths between start and end nodes) would take:
N^2 cells total
2N cells already occupied with start and end nodes
N^2 - 2N cells for which the color has yet to be determined
N colours available.
N^(N^2 - 2N) possible combinations.
So,
For N=5, this means 5^15 = 30517578125 combinations.
For N=6, this means 6^24 = 4738381338321616896 combinations.
In other words, the number of possible combinations is pretty high to start with, but also grows ridiculously fast once you start making the board larger.
Constraining the number of cells per color
Obviously, we should try to reduce the number of configurations as much as possible. One way of doing that is to consider the minimum distance ("dMin") between each color's start and end cell - we know that there should at least be this much cells with that color. Calculating the minimum distance can be done with a simple flood fill or Dijkstra's algorithm.
(N.B. Note that this entire next section only discusses the number of cells, but does not say anything about their locations)
In your example, this means (not counting the start and end cells)
dMin(orange) = 1
dMin(red) = 1
dMin(green) = 5
dMin(yellow) = 3
dMin(blue) = 5
This means that, of the 15 cells for which the color has yet to be determined, there have to be at least 1 orange, 1 red, 5 green, 3 yellow and 5 blue cells, also making a total of 15 cells.
For this particular example this means that connecting each color's start and end cell by (one of) the shortest paths fills the entire board - i.e. after filling the board with the shortest paths no uncoloured cells remain. (This should be considered "luck", not every starting configuration of the board will cause this to happen).
Usually, after this step, we have a number of cells that can be freely coloured, let's call this number U. For N=5,
U = 15 - (dMin(orange) + dMin(red) + dMin(green) + dMin(yellow) + dMin(blue))
Because these cells can take any colour, we can also determine the maximum number of cells that can have a particular colour:
dMax(orange) = dMin(orange) + U
dMax(red) = dMin(red) + U
dMax(green) = dMin(green) + U
dMax(yellow) = dMin(yellow) + U
dMax(blue) = dMin(blue) + U
(In this particular example, U=0, so the minimum number of cells per colour is also the maximum).
Path-finding using the distance constraints
If we were to brute force enumerate all possible combinations using these color constraints, we would have a lot less combinations to worry about. More specifically, in this particular example we would have:
15! / (1! * 1! * 5! * 3! * 5!)
= 1307674368000 / 86400
= 15135120 combinations left, about a factor 2000 less.
However, this still doesn't give us the actual paths. so a better idea would be to a backtracking search, where we process each colour in turn and attempt to find all paths that:
doesn't cross an already coloured cell
Is not shorter than dMin(colour) and not longer than dMax(colour).
The second criteria will reduce the number of paths reported per colour, which causes the total number of paths to be tried to be greatly reduced (due to the combinatorial effect).
In pseudo-code:
function SolveLevel(initialBoard of size NxN)
{
foreach(colour on initialBoard)
{
Find startCell(colour) and endCell(colour)
minDistance(colour) = Length(ShortestPath(initialBoard, startCell(colour), endCell(colour)))
}
//Determine the number of uncoloured cells remaining after all shortest paths have been applied.
U = N^(N^2 - 2N) - (Sum of all minDistances)
firstColour = GetFirstColour(initialBoard)
ExplorePathsForColour(
initialBoard,
firstColour,
startCell(firstColour),
endCell(firstColour),
minDistance(firstColour),
U)
}
}
function ExplorePathsForColour(board, colour, startCell, endCell, minDistance, nrOfUncolouredCells)
{
maxDistance = minDistance + nrOfUncolouredCells
paths = FindAllPaths(board, colour, startCell, endCell, minDistance, maxDistance)
foreach(path in paths)
{
//Render all cells in 'path' on a copy of the board
boardCopy = Copy(board)
boardCopy = ApplyPath(boardCopy, path)
uRemaining = nrOfUncolouredCells - (Length(path) - minDistance)
//Recursively explore all paths for the next colour.
nextColour = NextColour(board, colour)
if(nextColour exists)
{
ExplorePathsForColour(
boardCopy,
nextColour,
startCell(nextColour),
endCell(nextColour),
minDistance(nextColour),
uRemaining)
}
else
{
//No more colours remaining to draw
if(uRemaining == 0)
{
//No more uncoloured cells remaining
Report boardCopy as a result
}
}
}
}
FindAllPaths
This only leaves FindAllPaths(board, colour, startCell, endCell, minDistance, maxDistance) to be implemented. The tricky thing here is that we're not searching for the shortest paths, but for any paths that fall in the range determined by minDistance and maxDistance. Hence, we can't just use Dijkstra's or A*, because they will only record the shortest path to each cell, not any possible detours.
One way of finding these paths would be to use a multi-dimensional array for the board, where
each cell is capable of storing multiple waypoints, and a waypoint is defined as the pair (previous waypoint, distance to origin). The previous waypoint is needed to be able to reconstruct the entire path once we've reached the destination, and the distance to origin
prevents us from exceeding the maxDistance.
Finding all paths can then be done by using a flood-fill like exploration from the startCell outwards, where for a given cell, each uncoloured or same-as-the-current-color-coloured neigbour is recursively explored (except the ones that form our current path to the origin) until we reach either the endCell or exceed the maxDistance.
An improvement on this strategy is that we don't explore from the startCell outwards to the endCell, but that we explore from both the startCell and endCell outwards in parallel, using Floor(maxDistance / 2) and Ceil(maxDistance / 2) as the respective maximum distances. For large values of maxDistance, this should reduce the number of explored cells from 2 * maxDistance^2 to maxDistance^2.
I think you'll want to do this in two steps. Step 1) find a set of non-intersecting paths that connect all your points, then 2) Grow/shift those paths to fill the entire board
My thoughts on Step 1 are to essentially perform Dijkstra like algorithm on all points simultaneously, growing together the paths. Similar to Dijkstra, I think you'll want to flood-fill out from each of your points, chosing which node to search next using some heuristic (My hunch says chosing points with the least degrees of freedom first, then by distance, might be a good one). Very differently from Dijkstra though I think we might be stuck with having to backtrack when we have multiple paths attempting to grow into the same node. (This could of course be fairly problematic on bigger maps, but might not be a big deal on small maps like the one you have above.)
You may also solve for some of the easier paths before you start the above algorithm, mainly to cut down on the number of backtracks needed. In specific, if you can make a trace between points along the edge of the board, you can guarantee that connecting those two points in that fashion would never interfere with other paths, so you can simply fill those in and take those guys out of the equation. You could then further iterate on this until all of these "quick and easy" paths are found by tracing along the borders of the board, or borders of existing paths. That algorithm would actually completely solve the above example board, but would undoubtedly fail elsewhere .. still, it would be very cheap to perform and would reduce your search time for the previous algorithm.
Alternatively
You could simply do a real Dijkstra's algorithm between each set of points, pathing out the closest points first (or trying them in some random orders a few times). This would probably work for a fair number of cases, and when it fails simply throw out the map and generate a new one.
Once you have Step 1 solved, Step 2 should be easier, though not necessarily trivial. To grow your paths, I think you'll want to grow your paths outward (so paths closest to walls first, growing towards the walls, then other inner paths outwards, etc.). To grow, I think you'll have two basic operations, flipping corners, and expanding into into adjacent pairs of empty squares.. that is to say, if you have a line like
.v<<.
v<...
v....
v....
First you'll want to flip the corners to fill in your edge spaces
v<<<.
v....
v....
v....
Then you'll want to expand into neighboring pairs of open space
v<<v.
v.^<.
v....
v....
v<<v.
>v^<.
v<...
v....
etc..
Note that what I've outlined wont guarantee a solution if one exists, but I think you should be able to find one most of the time if one exists, and then in the cases where the map has no solution, or the algorithm fails to find one, just throw out the map and try a different one :)
You have two choices:
Write a custom solver
Brute force it.
I used option (2) to generate Boggle type boards and it is VERY successful. If you go with Option (2), this is how you do it:
Tools needed:
Write a A* solver.
Write a random board creator
To solve:
Generate a random board consisting of only endpoints
while board is not solved:
get two endpoints closest to each other that are not yet solved
run A* to generate path
update board so next A* knows new board layout with new path marked as un-traversable.
At exit of loop, check success/fail (is whole board used/etc) and run again if needed
The A* on a 10x10 should run in hundredths of a second. You can probably solve 1k+ boards/second. So a 10 second run should get you several 'usable' boards.
Bonus points:
When generating levels for a IAP (in app purchase) level pack, remember to check for mirrors/rotations/reflections/etc so you don't have one board a copy of another (which is just lame).
Come up with a metric that will figure out if two boards are 'similar' and if so, ditch one of them.
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.