Analysing and combining partial problem solutions for properly informed heuristic generations

Computer Engineering and Intelligent Systems www.iiste.org
ISSN 2222-1719 (Paper) ISSN 2222-2863 (Online)
Vol 3, No.11, 2012

Analysing and Combining Partial Problem Solutions for
Properly Informed Heuristic Generations
1
Kee-cheol Lee * and Han-gyoo Kim
Computer Engineering Department, Hongik University, Seoul, Korea 121-791
* E-mail of the corresponding author: kclee1@hongik.ac.kr
Abstract
Finding an optimal path to the fixed goal state of a problem instance lying in an enormous search space
may be described in the framework of the conventional A* algorithm. However, the estimated distance to
the goal state, so called h_value, must be generated by an admissible heuristic such that it is not larger than
but still as close as the unknown real distance to the goal. In this paper, we suggest a method of generating
a heuristic with that property. After analyzing a number of devised partial problems, some are selected to be
combined to produce a properly informed heuristic. In solving a complex problem with a fixed goal, some
depth of fixed backward states is pre-stored. Those static backward states are also used for partial problem
backward searches. For a given problem instance, the forward search is first performed for each of its
partial problem. The dynamically generated space is combined with the static search space to produce the
temporary search space, which is used to aid in the generation of each state heuristic for the course of
problem solving. A novel method of constructing the temporary search space for each partial problem is
suggested, in which each forward state found in the static backward space is back-propagated and
propagated in the forward space. To show the effectiveness of our method, it has been massively
experimented for instances of Rubik’s cube problem of some difficulty whose search space of states
reachable from any given start state is known to cover 43*1018 states, the number of which even an 64-bit
unsigned integer cannot hold.
Keywords: A*, admissible heuristic, partial problems, dynamic forward search, static backward search,
Rubik’s cube

1. General Framework for Solving a Complex Problem
Most complex problems are technically classified as NP-hard problems, which mean that as the size of a
problem instance gets larger, its optimal solution gets virtually impossible to obtain. That is why
heuristic-based knowledge is generally used for generating a good solution instead of the optimal one.
However, to get the optimal (or even near optimal) one for a somewhat larger size of a problem instance,
the framework of A* algorithm may be tried. Figure 1 describes the A* algorithm (Nilsson et al. 1968;
Luger 2008) modified to utilize the pre-stored static backward space of states. Two auxiliary functions are
additionally used in the algorithm.

void State::set_h() { /* heuristic for calculating h_value */ }
unsigned int State::get_f() { return g_value + h_value; }
bool Modified_A* {
priority_queue<State> OPEN;
set<State> CLOSED; // CLOSED is a set of states
START.g_value = 0; START.set_h(); OPEN.push(START); // START into OPEN
while ( true ) {
if (OPEN.isempty()) return false; // no solution exists
State P = OPEN.top(); OPEN.pop(); //get state with max f
if (P in pre-stored static backward search space) {

1
This work is supported in part by Hongik University Research Fund 2012.

1

Vol 3, No.11, 2012

print path from START through P to GOAL; return true; }
for (each child C of P) {
C.g_value = P.g_value + 1; C.set_h();
if (C exists as oldC in OPEN or CLOSED)
{ if (C.get_f()>oldC.get_f()) { delete(oldC); OPEN.push(C); } }
else OPEN.push(C);
} // end of for each child …
CLOSED.delete(P);
} // end of while ( true )
} // end of Modified_A*
Figure 1. General Framework of the Modified A* algorithm
Let START denote the state of the given instance and GOAL be the static final state. In this version, GOAL
is one of the states in the pre-stored static backward space, and does not explicitly appear in Figure 1. Each
state has g_value, the number of steps from START to it, and h_value, the number of estimated steps from
it to GOAL. (Those two values may be differently defined in terms of moving costs instead of the number
of steps. However, without loss of generality we assume the latter.) The sum of the two values, f value
(returned by the function get_f()) is used to guide OPEN, the max priority queue holding the states ready to
expand. If the h_value of a state could always be set to its unknown actual number of steps (called
h*_value), A* algorithm would reduce to an ideal analytic solution. The main issue here is how to derive the
admissible (or nearly admissible) heuristic, i.e. function set_h(), such that its value is as large as possible
but still not larger than h*_value. A static pattern database (Korf 1997) may help, but we suggest a more
complicated method.

2. Generating Properly Informed Admissible Heuristics
We assume the problem has the fixed GOAL state and the static backward search space of states of some
depth has been pre-computed, which is just a one-time job. This may be classified as a method which
generates and combines some partial problem solutions (Lee & Kim 2012).

2.1 Preliminary procedure for the problem domain
For the given problem domain (e.g., Rubik’s cube), the following preliminary procedure must be done to
decide which partial problems must be used, and to prepare the corresponding static spaces.

2.1.1 Construction of static backward search space
Make a space BSS of states reachable (in the breadth first way) from GOAL to be used for backward
searches. Its depth, d_back, may be a design parameter and can be decided considering the disk space
reserved for storing static backward states.

2.1.2 Design of partial problems and their indexing schemes
Generate partial problems of the given problem. Solve them in the framework of A* for some (say 50)
random problems. The partial problems must be small enough to generate their optimal solutions fast, but
big enough to generate large h_values usable for solving the original problem. The different hashing
scheme suitable for each partial problem should be designed to access the same pre-stored static search
space. Its details will not be discussed in this paper

2.1.3 Selection of partial problems and construction of BSS_PARTIAL(i), the static backward part of
SS_PARTIAL(i)

2

Vol 3, No.11, 2012

Select some partial problems whose max h_values are large enough. Let’s call the backward search space
for the i-th selected partial problem BSS_PARTIAL(i). The new heuristic is defined to be the maximum of
the h_values of all selected partial problems.

2.1.4 Design of the proper depth of the dynamic forward search space
Decide the proper forward depth, d_for, by considering the max h_value found from partial problems
subtracted by the pre-stored depth of static backward search space.

2.2 Procedure for a given instance
For each new problem instance, the following procedure can be used.

2.2.1 Construction of FSS_PARTIAL(i), forward part of SS_PARTIAL(i)
We already have BSS_PARTIAL(i), the static part of SS_PARTIAL(i). Now the dynamic forward part of
FSS_PARTIAL(i) is built in this stage. While building FSS_PARTIAL(i), any newly generated state for the
i-th partial problem is tested if it already exists in BSS_PARTIAL(i). If that is the case, the known h_value of
the old one is the new h_value of the new one, and it is back-propagated back to START in the forward
search space. In addition, after FSS_PARTIAL(i) is temporarily constructed, all h_values of the states from
START to the final depth of the forward space must be propagated such that the h_values of any adjacent
states cannot be different by more than 1. The algorithm which reflects this is as follows.

for (int i = 0; i < #selected_partial_problems; i++) {
// make a dynamic forward search space FSS_PARTIAL(i)
START.h_value = d_for+d_back; START.g_value = 0; // h set to max value; g set to #steps from START
vector<State> V; V.addrear(START); // V is used as a big array ; start form START
ind = 0; State Cur, C, P;
while ( ind < V.size() ) {
CUR = V[ind++];
if (CUR.g_value < d_for) // init & push each child except for deepest depth
for (each child C of CUR) { C.h_value = CUR.h_value–1; C.g_value = CUR.g_value+1; V.addrear(C); }
if (CUR already exists in BSS_PARTIAL(i)) {
CUR.h_value = the depth of the stored state;
C = CUR;
// back-propagate the newly acquired h_value
while (C != START) {
Let P be the parent state of C;
if (P.h_value <= (C.h_value + 1)) break; // no more back-propagation
P.h_value=C.h_value+1; // lower P’s h_value
C = P; // for one step back
} // end of while (C != START) {
} // end of if (CUR…
} // end of while (ind < V.size() ) {
// Now propagation begins
for (int ind = 1; ind < V.size(); ind++) {
// start from level 1, not level 0
C = V[ind]; Let P be the parent state of C;
if (C.h_value > (P.h_value+1)) C.h_value = P.h_value + 1; // lower C’s h_value

3

Vol 3, No.11, 2012

} // end of for (int ind…
Store vector V as FSS_PARTIAL(i)
}

2.2.2 h_value calculating function, set_h
Please note that the above procedure is done once for a given problem instance. Those spaces for partial
problem instances are used throughout the problem solving. The heuristic-calculating function utilizes those
fixed spaces. For each state of a problem space, the largest of all partial problem h_values is the wanted
h_value. That value and g_value, the number of steps (i.e., depths) from START to the current state, are
added to produce the f_value of a state. The OPEN priority queue, which contains the states ready to
expand, is designed such that its top function returns the state with the largest f_value.

void State::set_h() {
h_value = 0;
for (int i = 0; i < #selected_partial problems; i++) {
if (this state in SS_PARTIAL(i)) new_h = its stored h_value;
else new_h = d_back + 1; // defaulted to 8 for d_back set to 7
if (new_h > h_value) h_value = new_h;
}
}

Figure 2. TFace and Tile Number Notation

3. The problem Space for Experiments
Just to clarify the validity of our method, we utilized a well-known game problem, Rubik’s cube, widely
considered to be the world’s best-selling toy (Indep. 2007). As of January 2009, 350 million cubes had been
sold worldwide (Adam 2009)(Jamieson 2009). A number of solutions have been developed which can solve
the cube in wee under 100 moves (Marshall 2005; Singmaster 1981), which are far from optimal, and we
are not interested in them.
Frey & Singmaster (1982) that the number of moves needed to solve any Rubik’s cube, given an ideal
algorithm, might be in “the low twenties”. Kunkle & Cooperman (2007) used computer search methods to
demonstrate that any Rubik’s cube can be solved in 26 moves or fewer. Rokicki (2008) lowered that
number to 22 moves, and in July 2010, researchers including Rokicki, with about 35 CPU-years of idle
computer time donated by Google, proved the so-called “God’s number” to be 20 (Flatley 2010). More

4

Vol 3, No.11, 2012

generally, it has been shown that an n x n x n Rubik’s cube can be solved optimally in the order of n2 / log(n)
moves (Dermaine et al. 2011).
In this paper, we tested the effectiveness of our suggested method in solving an optimal or near optimal
solution of a given Rubik’s cube of some difficulty in 20 or less steps.

3.1 The problem space
The Rubik’s cube has 6 faces each of which has 9 tiles. Because center tiles are fixed during any move,
every cube state can be defined by 48 tiles as in Fig. 2. For the puzzle to be solved, each face must be
returned to consisting of one color(Dempsey 1988).
The 48 tiles can be divided into two groups, 8 corner cubies of three tiles and 12 edge cubies of two tiles.
Corner (Edge) cubies move only to corner (edge) cubie positions. The entire space is also known to consist
of 12 separate but isomorphic sub-graphs with no legal moves between them (Korf 1997). Therefore, the
total number of states reachable from a given state is (12!*212)*(8!*38)/12 = 43,252,003,274,489,856,000,
which is greater than the max number even an unsigned 64-bit integer can hold (i.e., 17.592*10e18).
Because we know the number of reachable states and God’s number is 20, the average branching factor b
from depth 0 to 20 can be calculated to be 9.536 by solving the equation of the sum of 21 terms of the
geometric sequence, i.e., 1 * (b21 – 1) / (b – 1) = 43.252 * 1018.
Table 1. Static backward and dynamic forward search spaces
pre-stored backward
dynamic forward search space
search space

depth
depth #states
#states

0 1 20 43,252,003,274,489,856,000

1 19 19 ~43e18

2 262 18 ~42e18

3 3,502 17 ~13e18

4 46,741 16 ~1.2e18

5 621,649 15 98,929,809,184,629,089

6 8,240,087 14 7,564,662,997,504,768

7 109,043,123 13 575,342,418,679,410

8 1,441,386,411 12 43,689,000,394,782

9 19,037,866,206 11 3,314,574,738,534

10 251,285,929,522 10 251,285,929,522

3.2 Storing the static space for backward search
The bi-directional search helps to reduce the search space. For example, Table 1 (Rokicki 2010) shows that
if 10-step forward search space and 10-step backward search space is used, we may deal with the space of
0.5e12 states, whereas 20-step forward search would require a space of 43*e18 states. That reduction ratio
corresponds to 1 versus one hundred million. Each row in the table shows the number of states generated
for problem solving. For example, backward space of depth 9 can be used with forward space of 11.
Because the GOAL state is fixed, we can pre-calculate the space of backward search and store it in disk (or

5

Vol 3, No.11, 2012

physical memory, depending on its size). Contrarily, the forward space is different for each problem
instance and must be dynamically generated. If we pre-store d_back-step backward states, the forward
search can be limited to the depth of 20-d_back, because we now know that the total depth can be safely
limited to 20, God’s number. Considering one state needs 40 bytes, we can deduce from Table 1 that the
pre-stored disk space is 0.3 Gigabytes for depth 6, 4.4 Gigabytes for depth 7, 57.7 Gigabytes for depth 8,
761.5 Gigabytes for depth 9, and 10.1 Terabytes for depth 10. This may be time-consuming, but it should
be noted that it should be done and stored once for the Rubik’s cube domain. In our experiments, the value
of d_back is set to 7, and moderate 4.4 Gigabytes are used for that.
Table 2. h_values calculated for partial problems of 50 random problem instances

a. b. c. d. e. 2-pairs 3-pairs 4-pairs

0&1 2&3 4&5 1&2 3&5 a&b a&c a&d a&e abc abd abe abcd abce

max 13 13 12 13 12 13 13 13 13 13 13 13 13 13

avg. 10.80 10.98 10.74 10.84 10.84 11.44 11.22 11.14 11.38 11.56 11.54 11.64 11.60 11.68

σ 1.00 0.91 0.84 0.78 0.78 0.83 0.73 0.85 0.69 0.70 0.75 0.62 0.69 0.61

4. Selection of Partial problems
Table 2 summarizes the results obtained by combining partial problems of two faces. Face numbers 0-5
denote Top, Back, Left, Front, Right, and Bottom faces. For the experimental purpose, fifty random cubes
generated by sufficiently many moves have been utilized. All the partial problems of two faces have been
solved in not more than 13 steps. Out of those 5 basic partial problems, some pairs of their h_values have
been combined. The last case of 3 pairs is a good choice in that its average h_value is sufficiently
high(11.64), with relatively low(0.62) standard deviation(σ), which indicates the stability of the method.
That happens to be the case of using 3 partial problems of faces (a) 0 & 1, (b) 2 & 3, and (c) 3 & 5,
implying that using 5 faces with one face overlapped may be better than using all of 6 faces.

5. Experimental Results
For the experimental purpose, a series of Rubik’s cube problem instances with their optimal lengths 8-14
were consistently used. We counted the number of states stored while running the main A* algorithm.
Table 3 summarizes the experimental results. The first experiment was done using a heuristic (called heu6
here) which is the number of misplaced tiles of all the edge cubies divided by 4, and whose max value is
limited to 6. We won’t delve into its details here. Other experiments are all based on our suggested method
with different max depths(d_for) of dynamic forward search space, set to 5-7. The table shows that as the
value of d_for grows, harder problems may be solved with a much smaller number of states stored. The
value of the static space depth, d_back, is currently set to 7 such that it requires just 4 Giga byte disk space
to hold pre-stored states, but practically we could raise it to even 10 because it requires 10.1 Tera byte disk,
which is acceptable in modern computers, and because its construction is just one time job. The value of the
dynamic search space depth, d_for, can be effectively raised as long as memory capacity allows.

6

Vol 3, No.11, 2012

Table 3. Experimental results with 7-step pre-stored backward states
(a) forw. heuristic=heu6 (b) forw. heuristic=myh(d_for=5)
states stored for partial problems forward states
# forward #
0-1 2-3 3-5 stored
steps states stored steps
46741+α 5-bfs inside

8 9 8 711 556 446 9

9 66 9 423 202 198 24

10 957 10 475 267 346 63

11 33,987 11 321 269 297 237

12 628,964 12 253 137 132 837

13 9,558,799 13 184 134 92 1,321,397

14 > 50M 14 203 125 99 > 50M

(c) forw. heuristic=myh(d_for=6) (d) forw. heuristic=myh(d_for=7)
states stored for partial states stored for partial
forward states forward states
# problems of faces # problems of faces
stored stored
steps 0-1 2-3 3-5 steps 0-1 2-3 3-5

621,649+α 6-bfs inside 8,240,087+α 7-bfs inside

8 5,363 4,152 3,534 9 8 48,900 42,709 53,760 9

9 3,914 2,371 2,108 93 9 41,226 27,290 25,593 84

10 5,069 2,921 3,733 120 10 52,007 34,389 41,757 510

11 3,897 2,971 3,215 504 11 43,741 34,120 37,265 *10,338

12 2,805 1,877 1,658 1,011 12 32,881 24,237 22,430 12,747

13 2,240 1,773 1,503 8,896 13 26,710 23,501 20,435 96,696

14 2,407 1,839 1,467 46,545,179 14 30,604 24,658 20,167 198,515

6. Conclusion
The optimal solution of a complex problem is very hard to obtain as the size of the problem instance grows.
Therefore its heuristic-based suboptimal solution is often tried. If its optimal or near optimal solution is
really necessary, it can be tried in the framework of A* algorithms. The admissible heuristic for A*, h_value,
is the guess of the actual number of steps, h*_value, from the current state to the goal state and must not
exceed it. The generation of admissible (or almost admissible) and sufficiently informed heuristic, h_value,
for each state is the most important, given a problem instance.
We suggested a bidirectional search paradigm with the static backward search space and the dynamic
forward search space which utilizes the heuristic value generated by combining each devised partial
problem instance heuristic. For a problem domain (e.g., Rubik’s cube), some partial problems are selected
and their hashing schemes are designed to maintain their own static backward spaces. For every problem
instance given, its dynamic forward space of states is generated to be combined with its own static space. If
any new state being generated in the forward space of a partial problem already exists in the corresponding
backward space, its old h_value is back-propagated to the start state, and propagated in the dynamic

7

Vol 3, No.11, 2012

forward search space. The newly combined space for each partial problem is used to calculate its own
h_value. The maximum of all the h_values recommended by each partial problem is used as the h_value of
each newly generated state of the given problem instance.
To show the effectiveness of our suggested method, it has been successfully experimented for a series of
Rubik’s cube problem instances of some difficulty.

References
Adams, W.L. (2009), “The Rubik’s Cube: A Puzzling Success”, TIME, Jan. ed.
Dempsey, M.W (1988), Growing up with science: The illustrated encyclopedia of invention, London:
Marshall Cavendish, 1245.
Dermaine, E., Dermaine, M.L., Eisenstat, S., Lubiw, A. & Winslow, A. (2011), “Algorithms for Solving
Rubik’s Cubes”, arTiv:1106.5736v.
Flatley, J.F. (2010), “Rubik’s cube solved in twenty moves, 35 years of CPU time”, Engadget, Aug. 9.
Frey, A. & Singmaster, D. (1982), Handbook of Cubik Math, Enslow Publishers.
Hart, P.E., Nilsson, N.J. & Raphael, B. (1968), “A Formal Basis for the Heuristic Determination of
Minimum Cost Paths”, IEEE Trans. on Science and Cybernetics, 2, 100-107.
Indep. (2007), “Rubik’s Cube 25 years on: crazy toys, crazy times”, The Independent(London), Aug. 16.
Jamieson, A. (2009), “Rubik’s Cube inventor is back with Rubik’s 360”, The Daily Telegraph, Jan.
Korf, R.E. (1997), “Finding Optimal Solutions to Rubik’s Cube Using Pattern Databases”, AAAI/IAAI,
700-705.
Kunkle, D. & Cooperman, C. (2007), “Twenty-Six Moves Suffice for Rubik’s Cube”, Proc. of the
International Symposium on Symbolic and Algebraic Computation(ISSAC '07), ACM Press.
Lee, K. & Kim, H. (2012), “Analyzing and Combining TF-IDF based Text Retrieval Systems”, GESTS Int.
Trans. On Computer Science and Engineering, 67(1), 41-52.
Luger, G. (2008), “Heuristic Search”, Ch. 4, Artificial Intelligence, 6 ed., Pearson.
Marshall, P. (2005), “The Ultimate solution to Rubik’s cube”, http://helm.lu/cube/MarshallPhilipp/.
Rokicki, T. (2008), “Twenty-Two Moves Suffice”, http://cubezzz.dyndns.org/drupal/?q=node/view/121,
Drupol, Aug.12.
Rokicki, T. (2010), “God’s Number is 20”, http://www.cube20.org/.
Singmaster, D. (1981), “Notes on Rubik’s Magic Cube”, Harmondsworth, Eng: Penguin Books.

Kee-cheol Lee He was born in Seoul, Korea on Feb. 21, 1955. He received a BS degree in electronic
engineering from Seoul national University in 1977, a MS degree in computer science from Korea
Advanced Institute of Science in 1979, and a Ph.D degree in electrical and computer engineering from
University of Wisconsin-Madison in 1987. Since March 1989, he has been on the faculty of computer
engineering department, Hongik University, Seoul, Korea, and currently he is a professor. His academic and
research interests cover the fields of artificial intelligence, machine learning, and information retrieval.

Han-gyoo Kim Prof. H. Kim was born in Seoul, Korea in 1959. He received B.S. degree in mechanical
engineering from Seoul National University in 1981, and his Ph. D in computer science from University of
California at Berkeley in 1994. Since August 1994, he has been on the faculty of computer engineering
department, Hongik University, Seoul, Korea. His areas of research include networked storage systems,
scalable information retrieval systems, and high speed large scale big data systems.

8

This academic article was published by The International Institute for Science,
Technology and Education (IISTE). The IISTE is a pioneer in the Open Access
Publishing service based in the U.S. and Europe. The aim of the institute is
Accelerating Global Knowledge Sharing.

More information about the publisher can be found in the IISTE’s homepage:
http://www.iiste.org

CALL FOR PAPERS

The IISTE is currently hosting more than 30 peer-reviewed academic journals and
collaborating with academic institutions around the world. There’s no deadline for
submission. Prospective authors of IISTE journals can find the submission
instruction on the following page: http://www.iiste.org/Journals/

The IISTE editorial team promises to the review and publish all the qualified
submissions in a fast manner. All the journals articles are available online to the
readers all over the world without financial, legal, or technical barriers other than
those inseparable from gaining access to the internet itself. Printed version of the
journals is also available upon request of readers and authors.

IISTE Knowledge Sharing Partners

EBSCO, Index Copernicus, Ulrich's Periodicals Directory, JournalTOCS, PKP Open
Archives Harvester, Bielefeld Academic Search Engine, Elektronische
Zeitschriftenbibliothek EZB, Open J-Gate, OCLC WorldCat, Universe Digtial
Library , NewJour, Google Scholar

Analysing and combining partial problem solutions for properly informed heuristic generations

More Related Content

What's hot (20)

Similar to Analysing and combining partial problem solutions for properly informed heuristic generations (20)

More from Alexander Decker (20)

Analysing and combining partial problem solutions for properly informed heuristic generations