Anonymous user
Talk:Solve a Hidato puzzle: Difference between revisions
m
→solved it using constraint programming
| (44 intermediate revisions by 7 users not shown) | |||
Line 68:
: I do have a version that solves even those awkward cases (using a more careful speculative execution model but still based on the solution in the task) but it's rather longer. I'll put it on a sub-page. –[[User:Dkf|Donal Fellows]] 08:24, 2 May 2012 (UTC)
:: Extended version failed on the "awkward" example with a stock Tcl8.6. Guess it's not recent enough? --[[User:Ledrug|Ledrug]] 00:07, 3 May 2012 (UTC)
::: Yes. It needs a feature ([http://www.tcl.tk/cgi-bin/tct/tip/397.html improved object cloning]) that's not yet made it into a beta release. –[[User:Dkf|Donal Fellows]] 07:50, 3 May 2012 (UTC)
== Any good general algorithm? ==
I added a larger example in the C code, which is meant to make brute force search miserable. The C code can solve it, but it takes quite long; I expect D to behave similarly. The Perl/Python/Tcl code should be able to solve it, if only taking forever, since they all use the same exhaustive search method (Tcl is forced into such a situation, but anyway). The Mathprog code gives up on it on my machine, rather quickly, which can be said to be the bright side. --[[User:Ledrug]]
:You do not indicate how glpk gives up on your machine, I am guessing numerical instability. Evil Case 2 makes the problem much larger. I have added examples of using Mathprog suitable for this large example. Note the model is the same, I just used a basis recommended for large multi indexed problems with a nested structure. In all three cases the solution then takes less than 1 second.--[[User:Nigel Galloway|Nigel Galloway]] 15:38, 4 May 2012 (UTC)
:: The earlier Mathprog code failed with something like "no possible solutions", without more explicit explanations. The current code still fails on the "evil 2" example, after about a gadzillion lines of instability warnings, then this:<lang>Error: unable to factorize the basis matrix (1)
Sorry, basis recovery procedure not implemented yet
glp_intopt: cannot solve LP relaxation</lang>
:: My glpsol version string is <code>GLPSOL: GLPK LP/MIP Solver, v4.45</code>, which might be why it doesn't have <code>--minisat</code>. Examples were run without it, not sure if that switch was crucial or not. --[[User:Ledrug|Ledrug]] 19:53, 4 May 2012 (UTC)
:::The no feasible solution error implies that the data file you created contained an error, as there is a feasible solution to a correct data file. The --minisat is needed to avoid the numerical instability with a correct data file.--[[User:Nigel Galloway|Nigel Galloway]] 13:26, 5 May 2012 (UTC)
:::For an explanation of minisat you may wish to see http://en.wikipedia.org/wiki/Tseitin-Transformation briefly summized:
::::The Tseitin Transformation is used to produce a boolean equation in conjunctive normal form (CNF) from a combinatorial logic circuit so that it may be solved by a SAT solver. The naive approach is to write the circuit as an equation, and use De Morgan's law and distribution. However, this can result in an exponential increase in equation size. The Tseitin Transformation outputs an equation whose size has grown linearly relative to the input circuit's.
:::Exponential increase is bad when you increase the problems size, so we avoid it --[[User:Nigel Galloway|Nigel Galloway]] 13:47, 5 May 2012 (UTC)
:::I have added an example in Ruby which shows that no time problem exists if the path length between hints is reasonable. The new C version could check the length of the path it is looking for and not adopt the new strategy if it is reasonably short, hence not slowing down normal puzzles.--[[User:Nigel Galloway|Nigel Galloway]] 13:26, 5 May 2012 (UTC)
:::: Even though I noted about slowdowns at the front of the C code, I'm not really all that concerned about it. The code uses a flood fill to check connectivity, which is O(n) because there's no backtracking, n being number of cells. For puzzles that are simple (i.e. brute force would have been more on the polynomial side without this check), it adds another polynomial term to run time, which could be relatively big but in absolute terms is not an issue (you can't see, but I'm making violent handwaving gesture here). The 3x3 example requires 8 tries at filling cells with or without the checks, so it's definitely faster without, but it won't be noticeable. For exponentially backtracking puzzles, the hope is it will kill off some (most?) long fruitless searches early (without check, the 50x3 example tries to fill values to cells 27962062 times before finding solution; with it, 85). I could do fancier checks, but not before I see a good example that demands it. --[[User:Ledrug|Ledrug]] 20:10, 5 May 2012 (UTC)
:::The following graph shows an example where if the path 1 2 5 6 etc is chosen, everything thinks it has reasonable connectivity, but they are kidding themselves. Graphs like this have efficient general solutions, they occur for instance in garbage collectors and glpk.--[[User:Nigel Galloway|Nigel Galloway]] 13:26, 5 May 2012 (UTC)
[[File:Snake2.PNG|centerHidato problem]]
<br clear=both>
Note that that example is trivial for a human. Does anyone have any clever way to deal with such cases? --[[User:Ledrug|Ledrug]] 00:59, 3 May 2012 (UTC)
: The second "sturdy" D version is able to solve the stress (but it's a little slower than the C version because it uses higher level data structures and no pointers, I'll try to speed it up a bit). I think the first "light" D entry breaks on the stress test, so it should be improved by someone.
:: The first simple D version now seems to work correctly.
I suspect that the reason people can solve that problem easily is that they switch strategies easily. Coming up with a way to characterize how to switch is the hard part. (The Tcl code is much more optimized for the case where it's dealing with path sections that are fairly short and which mostly stretch between end points; this is a useful strategy with the majority of Hidato puzzles published in the press, but is far less useful with these evil edge cases.) –[[User:Dkf|Donal Fellows]] 08:07, 3 May 2012 (UTC)
: Heeey, who you callin' evil?
: I was curious if anyone has a good heuristic method that's reasonably simple and works for most cases, like the one used in knight's tour, but I guess it's a bit asking for much. It's just that a task with only an exponential brute force general solution is somewhat disappointing -- oh well. --[[User:Ledrug|Ledrug]] 11:35, 3 May 2012 (UTC)
:: “Evil” is actually quite complimentary.
:: The reason this is [[wp:NP-complete|hard]] is that it is really the finding of a [[wp:Hamiltonian path|Hamiltonian path]] (on a graph with a bound on the number of links per node); you just add in a (possibly non-planar) link between start and finish to see that this must be the case. I don't really feel like putting lots of effort into cracking very hard problems like this other than by simple brute force. Moreover, because it ''is'' finding a Hamiltonian path, there must be cases which are not easy for either computers or humans to solve. (Do I know what they are? No, but an easy solution to arbitrary Hidato would have tremendous application to many other problems like cryptography.)
:: (I'd have written about this earlier, but there's some kind of weird proxy stopping me from posting to RC from home.) –[[User:Dkf|Donal Fellows]] 16:33, 6 May 2012 (UTC)
::: I'm not sure Hidato is really equiv to a general Hamiltioian path problem -- is it known to be NP-complete? Hidato is much more constrained in that it has a unique solution requirement; its nodes can have only up to eight neighbors; and more importantly, spatial proximity between nodes are tightly related to the connectivity between them. Although I do agree that if a problem isn't mathematically well understood (as in, having a known effective algorithm), it's generally not worth putting too much effort into it. --[[User:Ledrug|Ledrug]] 17:49, 6 May 2012 (UTC)
:::: I'm also not ''sure'' that is HP, but the suspicion remains. I couldn't find anything in WP on requirements on connectivity for graphs doing HP; maybe there is something, but I couldn't find it. (Graph theory isn't my specialty, not at all.) –[[User:Dkf|Donal Fellows]] 08:55, 7 May 2012 (UTC)
== Short C version ==
I have a problem with this line of the short C version, maybe there's a small bug:
<lang c>while (!isspace(*s)) s++;</lang>
To show it replace that line with:
<lang c>while (!isspace(*s)) { printf("%u\n", s); s++; }</lang>
: Right. I though <code>isspace(0)</code> is true; it's not. --[[User:Ledrug|Ledrug]] 07:44, 3 May 2012 (UTC)
== The Snake in the Grass - A case for an Orphaned Cell Supervisor ==
The reason that Evil Case 2 is easy for people is that (most) are capable of learning.
The 'complex applet (should be servlet)in an external website' displays these puzzles on an interesting background, and finally animates a successful solution. Evil Case 2 should be a snake in the grass with a suitably caught rodent having it final journey as the snake's lunch for the successful animation. It should only be attempted on a VERY wide screen monitor.
[[File:Snake.PNG|center|Hidato problem]]
This snake problem comes down to 24 decisions similar to the one above. Suppose the path 1 3 4 5 is attempted. There is then no way back to 2. If your solution can not detect this until it is attempting to use 73 and the only cell you can reach already contains 74 the algorithm is going to be slow.
Obviously most people will detect this on the first corner and realize that every corner is the same.
For the computer cell 2 must complain when it becomes orphaned and prevent that selection asap.
--[[User:Nigel Galloway|Nigel Galloway]] 15:36, 4 May 2012 (UTC)
== On the application of Warnsdorff to Hidato ==
This task's origional premise was that the Knights Tour and Hidato are the same problem requiring only a change of adjacency list.
Normal Hidato are simpler than the Knight's Tour. But considering that Mathprog can solve the Knight's Tour without being commited to Intensive Care for emergency minisat due to induced numerical instability, we must conclude that it is possible to create Hidato's more difficult than Knight's Tours.
If they are the same problem then that which is truth for Knigh't Tour is truth for Hidato. Therfore it seems appropriate to Warnsorff Hidato.
Evil1 and the task's example indicate this does little harm (if only C programmers trained as doctors!). And now we snake it in 16 hunredths of a second.
Finally can we also do the Knight's Tour?
: Warnsdorf is truth for Knight's Tour, but it's not truth for Hidato, hence they are not the same problem (your logic). Warnsdorff this:
1 0 -1 -1 -1 0 0 -1 -1 -1 0 0 -1 -1 -1 0 0 -1 -1 -1 0 0 -1 -1 -1 0 0 -1 -1 -1 0 0 -1 -1 -1 0 0 -1 -1 -1 0 0 -1 -1 -1 82
-1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1
-1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1 0 -1 0 -1 -1
0 0 0 -1 -1 0 0 0 -1 -1 0 0 0 -1 -1 0 0 0 -1 -1 0 0 0 -1 -1 0 0 0 -1 -1 0 0 0 -1 -1 0 0 0 -1 -1 0 0 0 -1 -1 -1
:--[[User:Ledrug|Ledrug]] 17:29, 6 May 2012 (UTC)
== On the algebra of Hidato and Knights's Tour ==
To clarify that the problem is mathmatically well understood lets consider Evil case 1. In the general case for
<pre>
a
b c d
e f g
</pre>
The problem is to find a solution for the following simultaneous equations:
<pre>
a1 .. g7 are binary variables.
a = a1*1 + a2*2 + a3*3 + a4*4 + a5*5 + a6*6 + a7*7
...
g = g1*1 + g2*2 + g3*3 + g4*4 + g5*5 + g6*6 + g7*7
a1 + a2 + a3 + a4 + a5 + a6 + a7 = 1
...
g1 + g2 + g3 + g4 + g5 + g6 + g7 = 1
a1 + b1 + c1 + d1 + e1 + f1 + g1 = 1
...
a7 + b7 + c7 + d7 + e7 + f7 + g7 = 1
b2 + c2 + d2 = a1
...
b7 + c7 + d7 = a6
a2 + c2 + e2 + f2 = b1
...
a7 + c7 + e7 + f7 = b6
a2 + b2 + d2 + e2 + f2 + g2 = c1
...
a7 + b7 + d7 + e7 + f7 + g7 = c6
a2 + c2 + f2 + g2 = d1
...
a7 + c7 + f7 + g7 = d6
b2 + c2 + f2 = e1
...
b7 + c7 + f7 = e6
b2 + c2 + d2 + e2 + g2 = f1
...
b7 + c7 + d7 + e7 + g7 = f6
c2 + d2 + f2 = g2
...
c7 + d7 + f7 = g6
</pre>
In the particular case:
<pre>
4
b 7 d
1 f g
</pre>
Which for this simple case can be resolved by inspection to:
<pre>
b = b2*2 + b3*3
d = d3*3 + d5*5
f = f2*2 + f6*6
g = g3*3 + g6*6
b3 + d3 = 1
b5 + d5 = 1
b2 + f2 = 1
b6 + d6 + f6 + g6 = 1
b3 + b5 = 1
d3 + d5 = 1
f2 + f6 = 1
g6 = 1
</pre>
Putting g6 into the model and resolving again solves the problem by hand.
It is apparent that the number of equations increases rapidly with problem size. A method exists which makes this increase linear see http://en.wikipedia.org/wiki/Tseitin-Transformation.
--[[User:Nigel Galloway|Nigel Galloway]] ([[User talk:Nigel Galloway|talk]]) 14:12, 30 December 2013 (UTC)
: if a1 = 0, the equation b2 + c2 + d2 = a1 means "Non of a's neighbors is 2" which is not necessarily true. so this solution does not work. --[[User:Ak|Ak]] ([[User talk:Ak|talk]]) 17:10, 21 March 2014 (UTC)
==extra credit suggestion==
How about adding an extra credit for this task: support Numbrix puzzles as well as Hidato puzzles. -- [[User:Gerard Schildberger|Gerard Schildberger]] ([[User talk:Gerard Schildberger|talk]]) 04:50, 18 December 2013 (UTC)
The REXX programming example only needed a couple of '''if''' statements to solve Numbrix puzzles in addition to Hidato puzzles. -- [[User:Gerard Schildberger|Gerard Schildberger]] ([[User talk:Gerard Schildberger|talk]]) 04:50, 18 December 2013 (UTC)
::Two points. First I think it would be better to add a related task Solve Numbrix Puzzles. I know what Numbrix is but only from external servlets and the press. Explaining it to them what be's satisfaction would further complicate this task. Secondly I had an extra credit for showing that the same logic could solve the Knight's Tour, but them what be made me remove it. The Knight's Tour is exactly the same as Hidato, only the Knight has a different view of which cells are adjacent. Numbrix has only 4 adjacent cells whereas Knight's Tour and Hidato have 8.--[[User:Nigel Galloway|Nigel Galloway]] ([[User talk:Nigel Galloway|talk]]) 13:35, 30 December 2013 (UTC)
::: Numbrix puzzles (in my thinking) are just a simpler Hidato puzzle, in the manner of having less possible moves (i.e., being more restrictive). But I have no qualms of someone creating a Rosetta Code task just for Numbrix puzzles. -- [[User:Gerard Schildberger|Gerard Schildberger]] ([[User talk:Gerard Schildberger|talk]]) 16:20, 30 December 2013 (UTC)
::: As regarding the Knights Tour: in fact, I used the same code in the REXX example, I just tweaked/optimized the code a bit for speed. -- [[User:Gerard Schildberger|Gerard Schildberger]] ([[User talk:Gerard Schildberger|talk]]) 16:20, 30 December 2013 (UTC)
== solved it using constraint programming ==
Behold: http://hidoku-solver.appspot.com/
| |||