SERESYE is a Rete based rules engine written in Erlang, descended directly from the Eresye project by Francesca Gangemi and Corrado Santoro. In the following article we will describe how to use the system.
As it is widely known, a rule-based system is composed by a knowledge base, which stores a set of facts representing the 'universe of discourse' of a given application, and a set of production rules, which are used to infer knowledge and/or reason about the knowledge. A rule is activated when one or more facts match the template(s) given in the rule declaration: in such a case, the body of the rule contains a code that is thus executed
In SERESYE, facts are expressed by means of Erlang tuples or records, while rules are written using standard Erlang function clauses, whose declaration reports, in the clause head, the facts or fact templates that have to be matched for the rule to be activated and executed.
For more information about SERESYE please refer to the paper docs directory.
For more information about rule-based inference engines and expert systems, you can refer to the book: S. Russell and P. Norvig. Artificial Intelligence: A Modern Approach/2E. Prentice Hall, 2003.
To write an AI application with SERESYE the following steps have to be performed:
-
Indentify your universe of discourse and determine the facts that have to be used to represent such a world;
-
Indentify the rules that you need and write them by using, e.g. first-order-logic predicates or even natural language;
-
Implement the system by writing your rules as Erlang function clauses, according to the modality required by SERESYE.
We will design a system able to derive new knowledge using some inference rules and starting from a small set; as a sample application, we chose the domain of relatives: we will start from some base concepts, such as parent, male and female, and then, by means of a proper set of rules, we will derive the concepts of mother, father, sister, brother, grandmother and grandfather.
According to the list above, we will first derive the facts that will be used to represent our concepts. Given the set of relationships above, they will be represented by means of the following facts:
| # | Concept | Fact / Erlang tuple |
| 1 | X is male | {male, X} |
| 2 | X is female | {female, X} |
| 3 | X is Y's parent | {parent, X, Y} |
| 4 | X is Y's mother | {mother, X, Y} |
| 5 | X is Y's father | {father, X, Y} |
| 6 | X is Y's sister | {sister, X, Y} |
| 7 | X is Y's brother | {brother, X, Y} |
| 8 | X is Y's grandmother | {grandmother, X, Y} |
| 9 | X is Y's grandfather | {grandfather, X, Y} |
Concepts 1, 2 and 3 will be used as a base to derive the other ones.
The rule to derive the concept of mother is quite straightforward:
if X is female and X is Y's parent then X is Y's mother.
From the point of view of SERESYE, since knowledge is stored in the knowledge base of the engine, the rule above is translated into the following one: *if the facts {female, X} and {parent, X, Y} are asserted in the knowledge base, then we assert the fact {mother, X, Y}.
The rule mother can be thus written as follows:
%%
%% if (X is female) and (X is Y's parent) then (X is Y's mother)
%%
mother (Engine, {female, X}, {parent, X, Y}) ->
seresye:assert (Engine, {mother, X, Y}).
This concept can be easily derived by means of the following rule:
%%
%% if (X is male) and (X is Y's parent) then (X is Y's father)
%%
father (Engine, {male, X}, {parent, X, Y}) ->
seresye:assert (Engine, {father, X, Y}).
This concept can be expressed by the following rule:
if Y and Z have the same parent and Z is female, then Z
is the Y's sister.
The SERESYE rule used to map this concept is:
%%
%% if (Y and Z have the same parent X) and (Z is female)
%% then (Z is Y's sister)
%%
sister (Engine, {parent, X, Y}, {parent, X, Z}, {female, Z}) when Y =/= Z ->
seresye:assert (Engine, {sister, Z, Y}).
Please note the guard, which is needed to ensure that when Y and Z are bound to the same value, the rule is not activated (indeed this is possible since the same fact can match both the first and second 'parent' pattern).
Given the previous one, this concept is now quite simple to implement:
%%
%% if (Y and Z have the same parent X) and (Z is male)
%% then (Z is Y's brother)
%%
brother (Engine, {parent, X, Y}, {parent, X, Z}, {male, Z}) when Y =/= Z ->
seresye:assert (Engine, {brother, Z, Y}).
The former concept can be expressed by means of the rule:
if X is Y's mother and Y is Z's parent, then X is Z's
grandmother.</u>* The latter concept is now obvious.
Both can be implemented using the following SERESYE rules:
%%
%% if (X is Y's mother) and (Y is Z's parent)
%% then (X is Z's grandmother)
%%
grandmother (Engine, {mother, X, Y}, {parent, Y, Z}) ->
seresye:assert (Engine, {grandmother, X, Z}).
%%
%% if (X is Y's father) and (Y is Z's parent)
%% then (X is Z's grandfather)
%%
grandfather (Engine, {father, X, Y}, {parent, Y, Z}) ->
seresye:assert (Engine, {grandfather, X, Z}).
After writing the rules, we need to:
- define the rules to seresye
- instantiate the engine;
- populate the knowledge base with a set of initial facts.
We define the rules to SERESYE by defined a 'rules' attribute at the start of the module.
%%%
%%% relatives.erl
%%%
-module (relatives).
-export([father/3, grandfather/3, grandmother/3,
mother/3, brother/4, sister/4, start/0]).
-rules([mother, father, brother, sister, grandfather,
grandmother]).
We continue on to instantiate the engine and popoulate the knowledge base in the function start below:
start () ->
application:start(seresye) % Only if it is not already started
seresye:start(relatives),
seresye:add_rules(relatives, ?MODULE)
seresye:assert(relatives,
[{male, bob}, {male, corrado}, {male, mark}, {male, caesar},
{female, alice}, {female, sara}, {female, jane}, {female, anna},
{parent, jane, bob}, {parent, corrado, bob},
{parent, jane, mark}, {parent, corrado, mark},
{parent, jane, alice}, {parent, corrado, alice},
{parent, bob, caesar}, {parent, bob, anna},
{parent, sara, casear}, {parent, sara, anna}]),
ok.
As the listing reports, creating a new SERESYE engine implies to call the function seresye:start/1, giving the name of the engine to be created
Then, we have to add the rules to the engine by using the function seresye:add_rule/2: it takes two arguments, the name of the engine and a tuple representing the function in the form {Module, FuncName}; obviously the function Module:FuncName must be exported. Function add_rule has to be called for each rule that has to be added; for this reason, the code above has an iteration over the list of rules written before.
Finally, we populate the inference engine with a set of sample facts by giving them, in a list, to the function seresye:assert/2. To test our rules, we considered the relationships in the Figure below and assert only the facts for male, female and parent.
The final complete code of our AI application is thus the following:
%%%
%%% relatives.erl
%%%
-module (relatives).
-export([father/3, grandfather/3, grandmother/3,
mother/3, brother/4, sister/4, start/0]).
%%
%% if (X is female) and (X is Y's parent) then (X is Y's mother)
%%
mother(Engine, {female, X}, {parent, X, Y}) ->
seresye:assert(Engine, {mother, X, Y}).
%%
%% if (X is male) and (X is Y's parent) then (X is Y's father)
%%
father(Engine, {male, X}, {parent, X, Y}) ->
seresye:assert(Engine, {father, X, Y}).
%%
%% if (Y and Z have the same parent X) and (Z is female)
%% then (Z is Y's sister)
%%
sister(Engine, {parent, X, Y}, {parent, X, Z}, {female, Z}) when Y =/= Z ->
seresye:assert(Engine, {sister, Z, Y}).
%%
%% if (Y and Z have the same parent X) and (Z is male)
%% then (Z is Y's brother)
%%
brother(Engine, {parent, X, Y}, {parent, X, Z}, {male, Z}) when Y =/= Z ->
seresye:assert(Engine, {brother, Z, Y}).
%%
%% if (X is Y's father) and (Y is Z's parent)
%% then (X is Z's grandfather)
%%
grandfather (Engine, {father, X, Y}, {parent, Y, Z}) ->
seresye:assert (Engine, {grandfather, X, Z}).
%%
%% if (X is Y's mother) and (Y is Z's parent)
%% then (X is Z's grandmother)
%%
grandmother(Engine, {mother, X, Y}, {parent, Y, Z}) ->
seresye:assert(Engine, {grandmother, X, Z}).
start () ->
application:start(seresye),
seresye:start (relatives),
seresye:add_rules(relatives, ?MODULE)
seresye:assert (relatives,
[{male, bob},
{male, corrado},
{male, mark},
{male, caesar},
{female, alice},
{female, sara},
{female, jane},
{female, anna},
{parent, jane, bob},
{parent, corrado, bob},
{parent, jane, mark},
{parent, corrado, mark},
{parent, jane, alice},
{parent, corrado, alice},
{parent, bob, caesar},
{parent, bob, anna},
{parent, sara, casear},
{parent, sara, anna}]),
ok.
Now it's time to test our application:
Erlang (BEAM) emulator version 5.5 [source] [async-threads:0] [hipe]
Eshell V5.5 (abort with ^G)
1> c(relatives).
{ok,relatives}
2> relatives:start().
ok
3>
Following the call to function relatives:start/0, the engine is created and populated; if no errors occurred, the rules should have been processed and the new facts derived. To check this, we can use the function seresye:get_kb/1, which returns the list of facts asserted into the knowledge base of a given engine:
4> seresye:get_kb(relatives).
[{brother,bob,mark},
{sister,alice,bob},
{sister,alice,mark},
{brother,bob,alice},
{brother,mark,alice},
{grandmother,jane,caesar},
{grandfather,corrado,caesar},
{grandmother,jane,anna},
{grandfather,corrado,anna},
{sister,anna,caesar},
{brother,caesar,anna},
{sister,anna,casear},
{mother,sara,anna},
{mother,sara,casear},
{parent,sara,anna},
{father,bob,anna},
{parent,sara,casear},
{father,bob,caesar},
{parent,bob,anna},
{father,corrado,alice},
{parent,bob,caesar},
{mother,jane,alice},
{parent,corrado,alice},
{father,corrado,mark},
{parent,jane,alice},
{mother,jane,mark},
{parent,corrado|...},
{brother|...},
{...}|...]
5>
The presence of facts representing concepts like father, sister, etc., proves that the rules seems to be working as expected.
We can however query the knowledge base using specific fact templates. For example, if we want to know who are Alice's brothers, we can use the function seresye:query_kb/2 as follows:
6> seresye:query_kb(relatives, {brother, '_', alice}).
[{brother,bob,alice},{brother,mark,alice}]
7>
The facts returned conform to the relationships depicted in the figure above, thus proving that the rules written are really working.
As the example shows, function seresye:query_kb/2 takes the engine name as the first argument, while, for the second parameter, a tuple has to be specified, representing the fact template to be matched; in such a tuple, the atom '_' plays the role of a wildcard. However, to specify a more complex matching, a fun can be used as a tuple element; this fun has to return a boolean value which indicates if the element matches the template. For example, to select both Alice's and Anna's brothers, we can use the following function call:
7> seresye:query_kb(relatives, {brother, '_', fun (X) -> (X == alice) or (X == anna) end}).
[{brother,bob,alice},{brother,mark,alice},{brother,caesar,anna}]
8>
This HowTo not only shows how to use the SERESYE engine to write an AI application, but also highlights the versatility of the Erlang language: the characteristics of functional and symbolic programming, together with the possibility of performing introspection of function declaration, can be successfully exploited for application domains which are completely new for Erlang but can surely be very interesting.