可以将文章内容翻译成中文,广告屏蔽插件可能会导致该功能失效(如失效，请关闭广告屏蔽插件后再试):
问题:
I am looking for an algorithm to detect redundant rules.
Rules have a fixed number of input parameters and each parameter has a distinct domain.
Consider three rule parameters Color, Material and Size:
- Color: Red, Green, Blue
- Material: Wood, Glass, Aluminium
- Size: Small, Medium, Large
Each rule can match on multiple values of a parameter or match on any value. The first rule that matches all parameters values is selected. There are no negation rules, but the domain is fixed, so a negation can be achieved by added all others.
+--------------------------------------------------++-----------------
| Rule Parameters || Rule Action
+----------------+------------------+--------------++-----------------
| Color | Material | Size || ==> Price
+----------------+------------------+--------------++-----------------
Rule 1 | Red | Wood, Glass | Large || $100
Rule 2 | Red | Aluminium | Large || $200
Rule 3 | Blue or Green | Wood | Medium || $300
Rule 4 | Red | ** Any ** | Large || $400 <== Redundant
Rule 5 | ** Any ** | ** Any ** | ** Any ** || $500
Rule 4 is redundant due to a combination of Rule 1 and 2. A redundant rule is a rule that will never be matched due to (a combination) of rules defined before that rule. The rule action is not evaluated in the redundancy check.
How to implement this efficiently (in Java)? It should support 1000 rules with 10 parameters with 100 values each. The rules, parameters and parameter values are read from a database (ie. they cannot be hard-coded).
The issue with efficiency is that there are 100^10 combinations of input parameters (each parameter has a domain of 100 values). The algorithm is needed in the rule editor gui, to support the user that is creating the rules. It should find all redundant rules within seconds.
GITHUB
I've created a repository to test proposed solutions: https://github.com/qlp/redundant-rules
Currently only a BDD implementation, which is failing with a problem of this size. Maybe my BDD implementation can be improved?
回答1:
Essentially, you're tasked with implementing a decision tree. Each level of the tree corresponds to each parameter, and at each level, there will be a node for each value that the parameter could hold. At the leaf level, you would store the rule action that was applicable.
For example, at the Color level, you'd have 3 nodes, Red, Blue, Green. Each of those would have 3 children (on the Material level) Wood, Glass, Aluminum. Each of those would have 3 children (Small, Medium, Large). You would construct this tree based on learning the parameters and values from the database.
Once the tree is constructed, you'd read the rules from the database, one at a time, and store the 'rule action' at each of the leaves (in this case, at the Size level). If there was a rule that wanted to apply at a leaf that already had an action, then there is an ambiguity about which rule would apply. For your problem you stated that you'd take the first rule that applied - so you would not change the rule stored.
If there was a rule, where for each of the leaves it had an action for there already was a defined action, then that rule would be 'redundant' according to your example.
回答2:
EDIT Completely rewritten due to the comments. (Note that this might actually be similar to what Khaled A Khunaifer wrote, but it has been created at the same time)
One possible approach could be to find a "logical" description of the rules. The rules have a quite regular form, namely always a conjunction of disjunctions. The rules can be written as
(Red) AND (Wood OR Glass) AND (Large)
(Red) AND (Aluminium) AND (Large)
...
Now, one could apply sophisticated optimizations and minimizations (like Quine McCluskey or any other form of minimization). However, I sketched a very simple implementation here. Rules can be "merged" when they only differ in one of the disjunctions. For example, the rules above could be merged to
(Red) AND (Wood OR Glass OR Aluminium) AND (Large)
Now, if a rule like
(Red) AND (Wood OR Glass OR Aluminium) AND (Medium)
was encountered, it could be merged with the previous one to
(Red) AND (Wood OR Glass OR Aluminium) AND (Large OR Medium)
If we then found a rule like
(Red) AND (Glass OR Aluminium) AND (Medium)
it could be marked as "redundant", because it is implied by the previous one.
To emphasize this again: This implementation is rather "hacky" and far from being "elegant", and there are certainly many possible improvements. But maybe it shows that the general idea is feasible, at least.
import java.util.ArrayList;
import java.util.Arrays;
import java.util.Collection;
import java.util.LinkedHashSet;
import java.util.List;
import java.util.Set;
public class RulesChecker
{
public static void main(String[] args)
{
List<Rule> rules = new ArrayList<Rule>();
List<Set<String>> domains = new ArrayList<Set<String>>();
domains.add(setOf("Red", "Green", "Blue"));
domains.add(setOf("Wood", "Glass", "Aluminium"));
domains.add(setOf("Small", "Medium", "Large"));
// Rule 1 | Red | Wood, Glass | Large || $100
// Rule 2 | Red | Aluminium | Large || $200
// Rule 3 | Blue or Green | Wood | Medium || $300
// Rule 4 | Red | ** Any ** | Large || $400 <== Redundant
// Rule 5 | ** Any ** | ** Any ** | ** Any ** || $500
rules.add(new Rule("$100", disOf("Red") , disOf("Wood", "Glass"), disOf("Large")));
rules.add(new Rule("$200", disOf("Red") , disOf("Aluminium") , disOf("Large")));
rules.add(new Rule("$300", disOf("Blue", "Green"), disOf("Wood") , disOf("Medium")));
rules.add(new Rule("$310", disOf("Blue") , disOf("Wood") , disOf("Medium")));
rules.add(new Rule("$350", disOf("Green") , disOf("Aluminium") , disOf("Medium")));
rules.add(new Rule("$400", disOf("Red") , disOf(domains.get(1)) , disOf("Large")));
rules.add(new Rule("$500", disOf(domains.get(0)) , disOf(domains.get(1)) , disOf(domains.get(2))));
System.out.println("Rules: ");
for (Rule rule : rules)
{
System.out.println(rule);
}
List<Rule> mergedRules = new ArrayList<Rule>();
mergedRules.add(rules.get(0));
for (int i=1; i<rules.size(); i++)
{
add(mergedRules, rules.get(i));
}
}
private static void add(List<Rule> mergedRules, Rule newRule)
{
for (int i=0; i<mergedRules.size(); i++)
{
Rule oldRule = mergedRules.get(i);
if (implies(oldRule, newRule))
{
System.out.println("Redundant "+newRule);
System.out.println(" due to "+oldRule);
return;
}
int mergeIndex = mergeIndex(oldRule, newRule);
if (mergeIndex != -1)
{
Rule mergedRule = merge(oldRule, newRule, mergeIndex);
mergedRules.set(i, mergedRule);
System.out.println("Merging "+oldRule);
System.out.println(" and "+newRule);
System.out.println(" gives "+mergedRule);
return;
}
}
mergedRules.add(newRule);
}
private static boolean implies(Rule oldRule, Rule newRule)
{
Conjunction c0 = oldRule.conjunction;
Conjunction c1 = newRule.conjunction;
List<Expression> es0 = new ArrayList<Expression>(c0.terms);
List<Expression> es1 = new ArrayList<Expression>(c1.terms);
for (int i=0; i<es0.size(); i++)
{
Disjunction d0 = (Disjunction) es0.get(i);
Disjunction d1 = (Disjunction) es1.get(i);
if (!d0.terms.containsAll(d1.terms))
{
return false;
}
}
return true;
}
private static Disjunction disOf(String ... ss)
{
return disOf(Arrays.asList(ss));
}
private static Disjunction disOf(Collection<String> ss)
{
List<Variable> list = new ArrayList<Variable>();
for (String s : ss)
{
list.add(new Variable(s));
}
return new Disjunction(list);
}
private static int mergeIndex(Rule r0, Rule r1)
{
Conjunction c0 = r0.conjunction;
Conjunction c1 = r1.conjunction;
List<Expression> es0 = new ArrayList<Expression>(c0.terms);
List<Expression> es1 = new ArrayList<Expression>(c1.terms);
int different = 0;
int mergeIndex = -1;
for (int i=0; i<es0.size(); i++)
{
Expression e0 = es0.get(i);
Expression e1 = es1.get(i);
if (!e0.equals(e1))
{
mergeIndex = i;
different++;
if (different > 1)
{
return -1;
}
}
}
return mergeIndex;
}
private static Rule merge(Rule r0, Rule r1, int index)
{
Conjunction c0 = r0.conjunction;
Conjunction c1 = r1.conjunction;
List<Expression> es0 = new ArrayList<Expression>(c0.terms);
List<Expression> es1 = new ArrayList<Expression>(c1.terms);
Set<Disjunction> rc = new LinkedHashSet<Disjunction>();
for (int i=0; i<es0.size(); i++)
{
Disjunction d0 = (Disjunction) es0.get(i);
Disjunction d1 = (Disjunction) es1.get(i);
if (i == index)
{
Set<Expression> merged = new LinkedHashSet<Expression>();
merged.addAll(d0.terms);
merged.addAll(d1.terms);
Disjunction d = new Disjunction(merged);
rc.add(d);
}
else
{
rc.add(d0);
}
}
return new Rule("TRUE", new Conjunction(rc));
}
private static Set<String> setOf(String ... s)
{
return new LinkedHashSet<String>(Arrays.asList(s));
}
static class Expression
{
}
static class Variable extends Expression
{
final String name;
Variable(String name)
{
this.name = name;
}
@Override
public String toString()
{
return name;
}
@Override
public int hashCode()
{
final int prime = 31;
int result = 1;
result = prime * result + ((name == null) ? 0 : name.hashCode());
return result;
}
@Override
public boolean equals(Object obj)
{
if (this == obj)
return true;
if (obj == null)
return false;
if (getClass() != obj.getClass())
return false;
Variable other = (Variable) obj;
if (name == null)
{
if (other.name != null)
return false;
}
else if (!name.equals(other.name))
return false;
return true;
}
}
static class Disjunction extends Expression
{
private final Set<Expression> terms;
Disjunction(Collection<? extends Expression> expressions)
{
this.terms = new LinkedHashSet<Expression>(expressions);
}
@Override
public String toString()
{
StringBuilder sb = new StringBuilder();
sb.append("(");
int n = 0;
for (Expression e : terms)
{
sb.append(e);
n++;
if (n < terms.size())
{
sb.append(" + ");
}
}
sb.append(")");
return sb.toString();
}
@Override
public int hashCode()
{
final int prime = 31;
int result = 1;
result = prime * result + ((terms == null) ? 0 : terms.hashCode());
return result;
}
@Override
public boolean equals(Object obj)
{
if (this == obj)
return true;
if (obj == null)
return false;
if (getClass() != obj.getClass())
return false;
Disjunction other = (Disjunction) obj;
if (terms == null)
{
if (other.terms != null)
return false;
}
else if (!terms.equals(other.terms))
return false;
return true;
}
}
static class Conjunction
{
private final Set<Expression> terms;
Conjunction(Collection<? extends Expression> expressions)
{
this.terms = new LinkedHashSet<Expression>(expressions);
}
@Override
public String toString()
{
StringBuilder sb = new StringBuilder();
sb.append("(");
int n = 0;
for (Expression e : terms)
{
sb.append(e);
n++;
if (n < terms.size())
{
sb.append(" * ");
}
}
sb.append(")");
return sb.toString();
}
@Override
public int hashCode()
{
final int prime = 31;
int result = 1;
result = prime * result + ((terms == null) ? 0 : terms.hashCode());
return result;
}
@Override
public boolean equals(Object obj)
{
if (this == obj)
return true;
if (obj == null)
return false;
if (getClass() != obj.getClass())
return false;
Conjunction other = (Conjunction) obj;
if (terms == null)
{
if (other.terms != null)
return false;
}
else if (!terms.equals(other.terms))
return false;
return true;
}
}
private static class Rule
{
Conjunction conjunction;
String action;
@SafeVarargs
Rule(String action, Disjunction ... disjunctions)
{
this.action = action;
this.conjunction = new Conjunction(Arrays.asList(disjunctions));
}
Rule(String action, Conjunction conjunction)
{
this.action = action;
this.conjunction = conjunction;
}
@Override
public String toString()
{
return conjunction+" -> "+action;
}
}
}
回答3:
One possibility is to use binary decision diagrams. These allow efficient manipulation of Boolean formulas. As mentioned by others, you can think of a rule as a binary formula: Rule 1 is (r OR (w AND g) OR l) and Rule 2 is (r AND a AND l), and to decide whether Rule 4 is redundant, we need to check whether (Rule4 AND (NOT (Rule1 OR Rule2 OR Rule3))) has a solution. Additionally, clauses like (NOT (w and g)) need to be used to disallow invalid inputs. There are many solvers available, and while there is no guarantee that running time or memory usage don't blow up, they typically work very well for real-world inputs.
回答4:
Each of your rules represents a boolean condition. It isn't clear how expressive your rule notation is (can I say, "not Wood"?), but Rule 1 appears to be
Color==Red and (Material==Wood or Material==Glass) and Size==Large
You also likely have some "domain constraints", C(k), presumably of the form:
Color==Red ==> Not(Color==Green)
where ==> is the logical implication operator.
What you appear to want to know, roughly, is if for some i and j, with R(i) meaning "ith Rule":
R(i) ==> R(j)
In practice, you need to know if:
R(i) and C(1) and C(2) and ... C(K) ==> R(j)
You can (and must) solve this by essentially doing boolean algebra.
This is actually relatively straightforward to do, e.g., treating the WHOLE equation (including the implication operator) as an actual symbolic formula and applying laws of boolean algebra.
Let's work your example algebraically (longish)
To start, for your example system of rules, the domain constraints C(i) are:
Color==Red ==> Not(Color == Green)
Color==Red ==> Not(Color == Blue)
Color==Blue ==> Not(Color == Green)
Color==Blue ==> Not(Color == Red)
Color==Green ==> Not(Color == Red)
Color==Green ==> Not(Color == Blue)
(Color==Red or Color==Blue or Color===Green) <==> true [if only these 3 colors exist]
and similarly for Materials (over Wood, Glass, Aluminum) and Size (Small, Medium, Large).
For your specific example, R(1) is:
Color==Red and (Material==Wood or Material==Glass) and Size==Large
and R(4) is:
Color==Red and (true) and Size==Large
What you want to know is if:
R(1) and <domainconstraints> ==> R(4)
That's a pretty huge formula, but that's really only a problem for the computer. In this particular case, we don't need the domain constraints (I, the oracle, speaketh) so I'm just going to leave it out to get rid of the hugeness:
R(1) and true ==> R(4)
or just
R(1) ==> R(4)
(Implementation hint: you can always try R(i) ==> R(j) first before trying R(i) and ==> R(j), because the first implies the second. This can be used to keep the size of the terms down if a redundancy is detected).
The "a ==> b" operator is boolean equivalent to " ~a or b" so you want to know if this formulat is true:
Not(R(1)) or R(4)
Plugging in the definitions of R(1) and R(4):
Not(Color==Red and (Material==Wood or Material==Glass) and Size==Large) or (Color==Red and (true) and Size==Large)
Using DeMorgan's law and simplifying out "and (true)":
Not(Color==Red) or Not( (Material==Wood or Material==Glass) ) or Not( Size==Large) or Color==Red and Size==Large
Apply DeMorgan's law a second time (we don't realy need this because Redness is going to drive the answer but we're not supposed to know that yet):
Not(Color==Red) or Not(Material==Wood) and Not(Material==Glass) or Not(Size==Large) or Color==Red and Size==Large
A fact:
a and b ==> a
so, for any b,
a and b or a == a
Using this with a being Not(Color==Red) and b being Size==Large:
Not(Color==Red) and Size==Large or Not(Color==Red) or or Not(Material==Wood) and Not(Material==Glass) or Not(Size==Large) or Color==Red and Size==Large
Now we group like terms:
Not(Color==Red) and Size==Large or Color==Red and Size==Large or Not(Color==Red) or or Not(Material==Wood) and Not(Material==Glass) or Not(Size==Large)
Combining terms having Size==Large:
( Not(Color==Red) or Color==Red) and Size==Large Not(Color==Red) or Not(Material==Wood) and Not(Material==Glass) or Not(Size==Large)
using a or ~ a == true:
( true ) and Size==Large or Not(Color==Red) or Not(Material==Wood) and Not(Material==Glass) or Not(Size==Large)
simplifying and grouping Size terms:
(Size==Large or Not(Size==Large)) or Not(Color==Red) or Not(Material==Wood) and Not(Material==Glass)
giving (whew)
(true) or ...
which produces true, so R(1) ==> R(4) and thus R(4) is redundant.
Implementation (or TL;DR)
Obviously you don't want to do this by hand. You can encode the formulas as boolean expressions (you've already done something like this to be able to store them in your system). What you want to know is if this statement as a whole is a tautology; for that, you could use Wang's rules of inference. That's a little clumsy to implement but doable.
Binary Decision Diagrams (awful name) are a way to encode "truth tables" for a formula. What you want is to decide if the truth table is always true for a formula. As another poster noted, you can get BDD packages. If you build the BDD for this implication equation, and it is "TRUE" everywhere (that is, you get a trivial BDD), then the implication is true and you have a redundancy.
The running time of checking this might get to be expensive. You can either put up with that expense, or simply put an upper bound on how much CPU you will give each attempt; either it produces an answer "no conflict", "conflict", or "timeout", with "timeout" being interpreted as "no conflict". You can then make this run as fast as you like, modulo sometimes not eliminating redundant rules.
Since your rules are independent declarations, if you have N rules, you'll need to do this test approximately N^2 times. (The good news, is that if you have already built up a trusted database of N rules, adding a new rule will only require N tests).
回答5:
There is extensive work in the field of pattern matching within strongly typed functional languages. The question of what it actually is has even been talked about here, on stackoverflow.
The publication page of Luc Maranget contains many articles on the topic in various settings, some of which served as a basis for the OCaml pattern matching algorithm.
His paper on compiling pattern matching to good decision trees could be a definitive source of inspiration for your problem. It also provides references for previous articles on the same topic. The basic principle is to describe patten matching sets as integer matrices, along with algebraic operations.
I also highly recommend trying ocaml, or any sml implementation, to get a feel of that central feature of the ML language family. It might be also interesting for you to check out scala, which implements similar functionalities.
回答6:
The number of rules in this problem is quite large, but the number of values is limited. Thats why I think a rule by rule evaluation will not give the quickest results.
If the rules are grouped by value, multiple rules can be evaluated together.
The algorithm I came up with can still explode, but I hope is less likely to do so.
Sample data:
Rule #: [Color (R,G,B),Material (W,G,A),Size (S,M,L)]
Rule 1: [[1,0,0],[1,1,0],[0,1,0]]
Rule 2: [[1,0,0],[0,0,1],[0,1,0]]
Rule 3: [[0,1,1],[1,0,0],[0,0,1]]
Rule 4: [[1,0,0],[1,1,1],[0,1,0]]
Rule 5: [[1,1,1],[1,1,1],[1,1,1]]
(Copied from Nuclearman)
Step 1: Split rules by value of parameter 1
Group the rules in a Set by a value for parameter 1.
Probably a heuristic is needed to pick the best first parameter.
For now I just pick them in order.
Result:
Rules in Set R: 1,2,4,5
Rules in Set G: 3,5
Rules in Set B: 3,5
Step 2: Merge Equal Sets
Set G en B contain the same rules at this level, so can be joined.
Result:
Rules in Set R: 1,2,4,5
Rules in Set GB: 3,5
Step 3: Split groups by value of parameter 2
For all the set the rule are now check on parameter 2.
For each different value the set is split into subsets.
The sets of step 2 are no longer relevant.
Result:
Rules in Set (R ,W): 1,4,5
Rules in Set (R ,G): 1,4,5
Rules in Set (R ,A): 2,4,5
Rules in Set (GB,W): 3,5
Rules in Set (GB,G): 5
Rules in Set (GB,A): 5
Step 4: Remove single rules
Rule 5 is now proven to be relevant. It singly catches (G,G,*), (B,G,*), (G,A,*) and (B,A,*).
Set (GB,G) and (GB,A) can be remove from the evaluation, as they have nothing to prove anymore.
Step 5: Merge Equal Sets
Result:
Rules in Set (R ,WG): 1,4,5
Rules in Set (R ,A ): 2,4,5
Rules in Set (GB,W ): 3,5
Step 3: Split groups by value for parameter 3
Result:
Rules in Set (R ,WG, S): 5
Rules in Set (R ,WG, M): 1,4,5
Rules in Set (R ,WG, L): 5
Rules in Set (R ,A ,S): 5
Rules in Set (R ,A ,M): 2,4,5
Rules in Set (R ,A ,L): 5
Rules in Set (GB,W ,S): 5
Rules in Set (GB,W ,M):5
Rules in Set (GB,W ,L): 3,5
Rule 5 is clearly proven to be useful, no counting rule sequence. 1, 2, 3 are proven, because there are the highest priority in there Set. Rule 4 is never prove and is redundant.
Now in pseudo-code:
void getProven(RuleList rulelist, RuleList provenlist, ParameterIterator i)
{
ValuesToRulelistMap subsets = new ValuesToRulelistMap();
Parameter parameter = i.next();
for (Rule rule : rulelist) {
Valuelist valuelist = rule.getValuesByParameter(parameter);
for (Value value : valueslist) {
subsets.addToListInMap(value, rule);
}
KeyList doneKeys = new KeyList();
for (RuleList subset : subsets.getRuleLists()) {
if (subset.length == 0) {
// No rule for this case!!!
} else
if (subset.length == 1) {
// Singly proven
provenlist.add(subset.getFirst());
} else
if (!i.hasNext()) {
// Proven by priority
provenlist.add(subset.getFirst());
} else
Key key = subset.combineRulesInListToKey()
if (!doneKeys.containts(key)) {
getProven(subset, provenlist, i);
doneKeys.add(key);
}
}
}
}
}
回答7:
First, let's take a look at your example, and instead of **ANY** we put all options
+---------------------------------++--------------
| Rule Parameters || Rule Action
+----------+-----------+----------++--------------
| Color | Material | Size || Price
+----------+-----------+----------++--------------
Rule 1 | R | W,G | L || $100
Rule 2 | R | A | L || $200
Rule 3 | B,G | W | M || $300
Rule 4 | R | A,W,G | L || $400
Rule 5 | R,B,G | A,W,G | S,M,L || $500
Next, we need to define how a Rule can be redundant, so here's my definition
A rule is redundant if it is equal to the merge of any k rules other than itself.
So, based on this definition (with the note that 1 <= k < n where n is number of rules) we can see that finding all redundant rules would take alot of time.
So, if we have n = 100 rules in our rule set, this brute-force will take around n^3 * n! which is Time ~ 100^3 * (1! + 2! + .. + 100!) = 10^6 * 9.4 * 10^158 = 9.333 * 10^164.
Here's my version of brute-force checking which takes O(n^3 * SUM { k! | k = 1..n } ):
boolean isRedundant (RuleList R, Rule d)
{
// R is the set of rules
// T is a copy of the set, without the rule to be checked
// d is the rule to be checked
T = new RuleList(R);
T.remove(d);
for (int i=0; i<T.size(); i++) // check single rules
if (T.get(j).equal(d))
return true;
for (int k=1; k<R.size(); k++) // 1 <= k < n
{
// merge every k-permutation to check
for (RuleList permutation : T.getPermutations(k))
{
Rule r = new Rule ();
for (Rule p : permutation)
r.merge(p);
if (r.equal(d))
return true;
}
}
return false;
}
Class Rule
class Rule
{
public boolean [3] color; // R,G,B
public boolean [3] material; // A,W,G
public boolean [3] size; // S,M,L
public Rule ()
{
color = { false, false, false };
material = { false, false, false };
size = { false, false, false };
}
public void merge (Rule r)
{
color[0] = and(color[0], r.color[0]);
color[1] = and(color[1], r.color[1]);
color[2] = and(color[2], r.color[2]);
material[0] = and(material[0], r.material[0]);
material[1] = and(material[1], r.material[1]);
material[2] = and(material[2], r.material[2]);
size[0] = and(size[0], r.size[0]);
size[1] = and(size[1], r.size[1]);
size[2] = and(size[2], r.size[2]);
}
public boolean equal (Rule r)
{
return (color[0]==r.color[0]
&& color[1]==r.color[1]
&& color[2]==r.color[2]
&& material[0]==r.material[0]
&& material[1]==r.material[1]
&& material[2]==r.material[2]
&& size[0]==r.size[0]
&& size[1]==r.size[1]
&& size[2]==r.size[2]);
}
public boolean and(boolean a, boolean b)
{
return (a && b);
}
}
Alternative Solution
Let's say that we have sets that represent each parameter, then we can reduce the number of rules to be checked from total number of rules, also this can optimize time complexity since we will see all members as one from that perspective, then go through a simple matching algorithm.
+---------------------------------++--------------
| Rule Parameters || Rule Action
+----------+-----------+----------++--------------
| Color | Material | Size || Price
+----------+-----------+----------++--------------
Rule 1 | R | W,G | L || $100
Rule 2 | R | A | L || $200
Rule 3 | B,G | W | M || $300
Rule 4 | R | A,W,G | L || $400
Rule 5 | R,B,G | A,W,G | S,M,L || $500
We have the following:
Color_R = { 1, 2, 4, 5 }
Color_B = { 3, 5 }
Color_G = { 3, 5 }
Material_A = { 2, 4, 5 }
Material_W = { 1, 3, 4, 5 }
Material_G = { 1, 4, 5 }
Size_S = { 5 }
Size_M = { 3, 5 }
Size_L = { 1, 2, 4, 5 }
Now, for every rule, we take its parameter sets, then match back without looking at the same rule ..
Example:
Rule_4 = { Color_R, Material_A, Material_W, Material_G, Size_L }
Color_R = { 1, 2, x, 5 }
Material_A = { 2, x, 5 }
Material_W = { 1, 3, x, 5 }
Material_G = { 1, x, 5 }
Size_L = { 1, 2, x, 5 }
We check for set of rules that match all of the memberships:
Matched = { 5, 1+2 }
Then we compare them with the selected rule, by set-difference looking for phi:
Rule_5 - Rule_4 = { Color_B, Color_G, Size_S, Size_M }
> Rule_5 does not match
(Rule_1 + Rule2) - Rule_4 = { }
> (Rule_1 + Rule2) match
Finally, we conclude:
(Rule_1 + Rule2) = Rule_4
> Rule_4 is redundant
