19. The Rete Algorithm

19.1. Disclaimer

The information in this Section is provided for the curious reader. An understanding of the Rete algorithm may be helpful in planning rule-based systems; an understanding of Jess's implementation probably will not. Feel free to skip this section and come back to it some other time. You should not take advantage of many of the Java classes mentioned in this section. They are internal implementation details and any Java code you write which uses them may well break each time a new version of Jess is released.

19.2. The Problem

Jess is a rule engine. In the simplest terms, this means that Jess's purpose it to continuously apply a set of if-then statements (rules) to a set of data (the working memory). You define the rules that make up your own particular rule-based system. Jess rules look something like this:
Jess> (defrule library-rule-1
    (book (name ?X) (status late) (borrower ?Y))
    (borrower (name ?Y) (address ?Z))
    (send-late-notice ?X ?Y ?Z))
This rule might be translated into pseudo-English as follows:
Library rule #1:
  a late book exists, with name X, borrowed by someone named Y
  that borrower's address is known to be Z
  send a late notice to Y at Z about the book X.
The book and borrower entities would be found on the working memory. The working memory is therefore a kind of database of bits of factual knowledge about the world. The attributes (called slots) that things like books and borrowers are allowed to have are defined in statements called deftemplates. Actions like send-late-notice can be defined in user-written functions in the Jess language (deffunctions) or in Java (Userfunctions) For more information about rule syntax refer to the chapter "Making your own Rules."

The typical rule-based program has a fixed set of rules while the working memory changes continuously. However, it is an empirical fact that, in most rule-based programs, much of the working memory is also fairly fixed from one rule operation to the next. Athough new facts arrive and old ones are removed at all times, the percentage of facts that change per unit time is generally fairly small. For this reason, the obvious implementation for the rule engine is very inefficient. This obvious implementation would be to keep a list of the rules and continuously cycle through the list, checking each one's left-hand-side (LHS) against the working memory and executing the right-hand-side (RHS) of any rules that apply. This is inefficient because most of the tests made on each cycle will have the same results as on the previous iteration. However, since the working memory is stable, most of the tests will be repeated. You might call this the rules finding facts approach and its computational complexity is of the order of O(RF^P), where R is the number of rules, P is the average number of patterns per rule LHS, and F is the number of facts on the working memory. This escalates dramatically as the number of patterns per rule increases.

19.3. The Solution

Jess instead uses a very efficient method known as the Rete (Latin for net) algorithm. The classic paper on the Rete algorithm ("Rete: A Fast Algorithm for the Many Pattern/ Many Object Pattern Match Problem", Charles L. Forgy, Artificial Intelligence 19 (1982), 17-37) became the basis for a whole generation of fast rule engines: OPS5, its descendant ART, CLIPS, and of course Jess. In the Rete algorithm, the inefficiency described above is alleviated (conceptually) by remembering past test results across iterations of the rule loop. Only new facts are tested against any rule LHSs. Additionally, as will be described below, new facts are tested against only the rule LHSs to which they are most likely to be relevant. As a result, the computational complexity per iteration drops to something more like O(RFP), or linear in the size of working memory. Our discussion of the Rete algorithm is necessarily brief. The interested reader is referred to the Forgy paper or to Giarratano and Riley, "Expert Systems: Principles and Programming", Second Edition, PWS Publishing (Boston, 1993) for a more detailed treatment.

The Rete algorithm is implemented by building a network of nodes, each of which represents one or more tests found on a rule LHS. Facts that are being added to or removed from the working memory are processed by this network of nodes. At the bottom of the network are nodes representing individual rules. When a set of facts filters all the way down to the bottom of the network, it has passed all the tests on the LHS of a particular rule and this set becomes an activation. The associated rule may have its RHS executed (fired) if the activation is not invalidated first by the removal of one or more facts from its activation set.

Within the network itself there are broadly two kinds of nodes: one-input and two-input nodes. One-input nodes perform tests on individual facts, while two-input nodes perform tests across facts and perform the grouping function. Subtypes of these two classes of node are also used and there are also auxilliary types such as the terminal nodes mentioned above.

An example is often useful at this point. The following rules:
(defrule example-2      (defrule example-3
    (x)                     (x)
    (y)                     (y)
    (z)                     => )
    => )
might be compiled into the following network:

The nodes marked x?, etc., test if a fact contains the given data, while the nodes marked + remember all facts and fire whenever they've received data from both their left and right inputs. To run the network, Jess presents new facts to each node at the top of the network as they added to the working memory. Each node takes input from the top and sends its output downwards. A single input node generally receives a fact from above, applies a test to it, and, if the test passes, sends the fact downward to the next node. If the test fails, the one-input nodes simply do nothing. The two-input nodes have to integrate facts from their left and right inputs, and in support of this, their behavior must be more complex. First, note that any facts that reach the top of a two-input node could potentially contribute to an activation: they pass all tests that can be applied to single facts. The two input nodes therefore must remember all facts that are presented to them, and attempt to group facts arriving on their left inputs with facts arriving on their right inputs to make up complete activation sets. A two-input node therefore has a left memory and a right memory. It is here in these memories that the inefficiency described above is avoided. A convenient distinction is to divide the network into two logical components: the single-input nodes comprise the pattern network, while the two-input nodes make up the join network.

19.4. Optimizations

There are two simple optimizations that can make Rete even better, The first is to share nodes in the pattern network. In the network above, there are five nodes across the top, although only three are distinct. We can modify the network to share these nodes across the two rules (the arrows coming out of the top of the x? and y? nodes are outputs):

But that's not all the redundancy in the original network. Now we see that there is one join node that is performing exactly the same function (integrating x,y pairs) in both rules, and we can share that also:

The pattern and join networks are collectively only half the size they were originally. This kind of sharing comes up very frequently in real systems and is a significant performance booster!

You can see the amount of sharing in a Jess network by using the watch compilations command. When a rule is compiled and this command has been previously executed, Jess prints a string of characters something like this, which is the actual output from compiling rule example-2, above:
  example-2: +1+1+1+1+1+1+2+2+t
Each time +1 appears in this string, a new one-input node is created. +2 indicates a new two-input node. Now watch what happens when we compile example-3:
  example-3: =1=1=1=1=2+t
Here we see that =1 is printed whenever a pre-existing one-input node is shared; =2 is printed when a two-input node is shared. +t represents the terminal nodes being created. (Note that the number of single-input nodes is larger than expected. Jess creates separate nodes that test for the head of each pattern and its length, rather than doing both of these tests in one node, as we implicitly do in our graphical example.) No new nodes are created for rule example-3. Jess shares existing nodes very efficiently in this case.

19.5. Implementation

Jess's Rete implementation is very literal. Different types of network nodes are represented by various subclasses of the Java class jess.Node: Node1, Node2, NodeNot2, NodeJoin, and NodeTerm. The Node1 class is further specialized because it contains a command member which causes it to act differently depending on the tests or functions it needs to perform. For example, there are specializations of Node1 which test the first field (called the head) of a fact, test the number of fields of a fact, test single slots within a fact, and compare two slots within a fact. There are further variations which participate in the handling of multifields and multislots. The Jess language code is parsed by the class jess.Jesp, while the actual network is assembled by code in the class jess.ReteCompiler. The execution of the network is handled by the class Rete. The jess.Main class itself is really just a small demonstration driver for the jess package, in which all of the interesting work is done.

The view command is a graphical viewer for the Rete network itself; I have used this as a debugging tool for Jess, but it may have educational value for others, and it may help you to design more efficient systems of rules in Jess. Issuing the view command after entering the rules example-2 and example-3 produces a very good facsimile of the drawing (although it correctly shows the larger number of one-input nodes). The various nodes are color-coded according to their roles in the network; Node1 nodes are red; Node2 nodes are green; NodeNot2 nodes are yellow; and Defrule nodes are blue. The orange node in the figure is a "right-to-left adapter" node; one of these is always used to connect the first pattern on a rule's LHS to the network. Passing the mouse over a node displays information about the node and the tests it contains; double-clicking on a node brings up a dialog box containing the same information (for join nodes, the memory contents are also displayed, while for Defrule nodes, a pretty-print representation of the rule is shown). See the description of the view function for important information before using it.

19.6. Efficiency of rule-based systems

Jess's rule engine uses an improved form of a well-known algorithm called Rete (Latin for "net") to match rules against the working memory. Jess is actually faster than some popular rule engines written in C, especially on large problems, where performance is dominated by algorithm quality.

Note that Rete is an algorithm that explicitly trades space for speed, so Jess' memory usage is not inconsiderable. Jess does contain some commands which will allow you to sacrifice some performance to decrease memory usage. Nevertheless, Jess' memory usage is not ridiculous, and moderate-sized programs will fit easily into Java's default 16M heap.

The single biggest determinant of Jess performance is the number of partial matches generated by your rules. You should always try to obey the following (sometimes contradictory) guidelines while writing your rules:

You can use the view command to find out how many partial matches your rules generate. See this chapter on How Jess Works for more details.

19.6.1. Sun's HotSpot Virtual Machine

Because Jess is a memory-intensive application, its performance is sensitive to the behavior of the Java garbage collector. Recent JVMs from Sun feature an advanced Java runtime called HotSpot which includes a flexible, configurable garbage collection subsystem. Excellent articles on GC performance tuning are available at Sun's web site. Although every Jess rule base is different, in general, Jess will benefit if the heap size and the object nursery size are each set larger than the default. For example, on my machine, Jess' performance on the Miranker manners benchmark with 90 guests is improved by 25% by increasing the initial heap size and nursery size to 32 and 16 megabytes, respectively, from their defaults of 16 meg and 640K. You can do this using
java -XX:NewSize=16m -Xms32m -Xmx32m jess.Main <scriptfile>
Note that the object nursery is a subset of the Java heap set aside for recently-allocated objects; the total heap size in this example is 32M, not 48M.