scriptNOP

A framework for Notification Oriented Paradigm (NOP[1]) implemented in TypeScript.

In the Notification Oriented Paradigm (NOP), there are the "factual and causal smart-entities named as Fact Base Elements (FBEs) and Rules that are related to another collaborative notifier smart-entities. Each FBE is related to Attributes and Methods, whereas each Rule to Premises-Conditions and Actions-Instigations. All these entities collaboratively carry out the inference process using Notifications, providing solutions to deficiencies of current paradigms" [1].

This implementation provides state-of-the-art features of NOP, in TypeScript, exploring the current limits of object orientation and imperative programming, parallel programming and concurrent programming. The implementation has no dependencies on other libraries and can be used in any TypeScript/JavaScript runtime or browsers.

Contents

• Sample application
• Defining Conditions
  • Condition with extensions
  • Combination of Conditions
• Rules
• Instructions to run this project
• Particularities of this implementation
  • Multithreaded and concurrent approach used
  • Comparison of existing concepts in the literature and new ones in relation to materializations of the NOP
  • Data typing
• References
• About

Sample application

This program contains an example of an application named "Target shooting". There is the main thread (state manager), where all the Fact Base Elements are, and there are the secondary threads where the Rules are.

main.ts (main thread):

import {
  App,
  delay,
  FactBaseElement,
} from "https://deno.land/x/script_nop/mod.ts";
App.init({
  numThreads: 1,
  extensionsURLs: [ //for URL to local file: new URL("./my_file.js", import.meta.url).href
    "https://deno.land/x/script_nop/src/extensions/deepEqual.ts",
  ],
  rulesURL: "https://deno.land/x/script_nop/example/rules.ts",
  onNotification: (n: any) => console.log(n), //Using history as a debugger
});
/*
  * Example: Target shooting aplication
  */
class Shooter extends FactBaseElement {
  constructor(fbeName: string) {
    super(fbeName);
  }
  shoot() {
    super.notify(
      "", //path
      {
        gun: {
          bullets: 5,
          pull_trigger: false,
        },
        target: false,
      },
    );
  }
}
const shooter1 = new Shooter("shooter1_name");
shooter1.shoot();

rules.ts (secondary threads):

import { Rule } from "https://deno.land/x/script_nop/mod.ts";
const rule1 = new Rule(
  {
    name: "rule1_name",
    condition: {
      left: {
        attribute: {
          fbe: "shooter1_name",
          attr: "gun.bullets",
        },
      },
      is: ">",
      right: {
        constant: 0,
      },
    },
    action: async (notifier: any, inferenceData: any) => { // Notifications from Conditions to Rules and Actions to Rules are accessible via inferenceData.
      console.log("loaded gun!!!");
      // get FBE Attribute value (using Notifications between Attributes and Actions).
      const bullets = await notifier.notifyAttrAndWaitReply({
        fbe: "shooter1_name",
        attr: "gun.bullets",
      });
      notifier.notifyFBE( // notify FBE
        "shooter1_name",
        "", //path
        {
          target: true,
        },
        "IGNORE_MISSING", //with "IGNORE_MISSING" mode, Attributes missing in "input" (in depth) will not be considered as excluded.
      );
      //Notifies a Rule to activate it (also depends on the Condition Notification of that Rule).
      notifier.notifyRule("rule2_name", { gun: "ok" });
    },
  },
);

For explicit Notification modes, if mode includes "IGNORE_MISSING", Attributes missing in "input" (in depth) will not be considered as excluded, if mode includes "FORCE" Notifications will be even if the values of these Notifications are not modified (there is no change in the status of the notifiable entity), and it is possible to combine the two modes ("IGNORE_MISSING_AND_FORCE"). FBEs do not necessarily need to be instantiated. If an FBE does not exist when being notified by another entity of the NOP core, its instance is created automatically. "IGNORE_MISSING" mode is only applicable for Notifications to FBEs

Defining Conditions

Conditions are implemented in a tree structure, easy for humans to understand. Note that the "." in Attributes is reserved in this implementation for path notation, and this implementation handles circular references. Conditions, Attributes values, and Notifications values ​​must be serializable to JSON. Conditions don't necessarily need to notify boolean values. However, these entities must notify a JavaScript true value for a Rule to activate it. Values like false, "", null, 0 and undefined will be false. The fact that Conditions notify each other with values ​​of any type allows the construction of chain operations. A Condition must have in its tree at least one notifying entity (and not just constants). Conditions can optionally accept parameters (such as fuzzy parameters[2]) and custom procedures like sum of a weighted input of a neuron[3] or an activation function like ReLU, and you can still combine it all at the same time. See examples of Conditions:

//------------------------- TYPES OF CONDITIONS ----------------------
//----------------WITH ONE ATTRIBUTE:
const c: Condition = {
  attribute: {
    fbe: "shooter1_name", //Fact Base Element name.
    attr: "target.person.age", //Attribute name - path notation
  },
};
//----------------WITH ONE CONSTANT:
const c: Condition = {
  constant: {
    name: "joe",
    age: 25,
  },
};
//---------------- CONDITION AS A PREMISE:
/*{
  left: Condition;
  is: string; //"==", ">", "<", etc.
  right: Condition;
}
*/
const c: Condition = {
  left: {
    attribute: {
      fbe: "shooter1_name",
      attr: "target.person.age",
    },
  },
  is: "==",
  right: {
    attribute: {
      fbe: "shooter1_name",
      attr: "target.person.age",
    },
  },
};
//---------------- CONDITION REFERENCING CONDITION FROM OTHER RULE:
const c: Condition = {
  ref: "rule2_name",
};
//----------------WITH OR, AND, XOR
const c: Condition = {
  and: [ //keys: "or", "and", "xor"
    //ARRAY of sub Conditions.
  ],
};
//----------------WITH CUSTOM PROCEDURE
const c: Condition = {
  op: "+", // "procedure name (registered extension) or operator (+, *, etc)",
  sub_conditions: [ //this vector is the input parameter of the procedure
    //ARRAY of sub Conditions.
  ],
};
//----------------WITH negation
const c: Condition = {
  not: c2, //c2 is one object of type Condition.
};
//----------------WITH OPTIONAL parameters:
//----------------FOR FUZZY LOGIC:
const c: Condition = {
  // ... (Condition body) ...
  min_threshold: { //These parameters can be any type of Condition
    constant: 0.2,
  },
  max_threshold: {
    constant: 0.8,
  },
};
const c: Condition = {
  // ... (Condition body) ...
  exactly: {
    constant: 0.5,
  },
};

Condition with extensions

There is also an extension interface for named procedures, which are used as customized operators in Conditions. These extensions are defined at the beginning of the main thread by the "extensionsURLs" parameter, but they can also be defined manually in the Rules file (in this case it must be before the Rules declaration):

DataMapChild.registerExtensions([customFunc2]);

How to use extensions:

/*
In Conditions:
custonFunc2 = procedure with name "custonFunc2", ex: export default function custonFunc2(items: any[]): any { ...
"items" is an array of result of Conditions ("sub_conditions" parameter).
*/
const c: Condition = {
  op: "customFunc2", //PROCEDURE NAME HERE, the "op" can also be operators like "+", "*", etc. Operators like "and" can be represented by their name in the language ("&&").
  sub_conditions: [ //"sub_conditions" only exists when the "op" attribute in a Condition is filled
    {
      left: {
        attribute: {
          fbe: "shooter1_name",
          attr: "gun.bullets",
        },
      },
      is: "==",
      right: {
        constant: 5,
      },
    },
  ],
};

Combination of Conditions

A combination of different types of Conditions together is possible. Example with simple logic, fuzzy logic and custom procedures:

const c: Condition = {
  or: [
    { // 'or' condition 1
      not: {
        op: "ReLU", //custom procedure name in Condition, input is sub_conditions Array
        sub_conditions: [
          {
            op: "sumOfWeights",
            sub_conditions: [
              {
                attribute: {
                  fbe: "layer1_name",
                  attr: "neurons.0", //paths with .N is valid for vectors
                },
              },
            ],
          },
        ],
      },
    },
    { // 'or' condition 2
      attribute: {
        fbe: "shooter1_name",
        attr: "gun.distance",
      },
      min_threshold: {
        constant: 0.2,
      },
    },
    { // 'or' condition 3
      left: {
        attribute: {
          fbe: "shooter1_name",
          attr: "gun.pull_trigger",
        },
      },
      is: "==",
      right: {
        constant: true,
      },
    },
  ],
};

In the application library, there is an extension procedure named "deepEqual", which checks in depth if two objects are the same, i.e. compares their parameters, subparameters and etc. It is possible for example an extension procedure that represents a sum of weighted weights of a neuron, it can also be combined with fuzzy logic for the activation threshold of the same.

Instructions to run this project

Basically you just need to clone the project and install the Deno runtime.

# clone project
git clone https://github.com/hviana/scriptNOP.git
# enter the project directory
cd scriptNOP
# install Deno (Mac, Linux)
curl -fsSL https://deno.land/install.sh | sh
# install Deno (Windows/PowerShell)
iwr https://deno.land/install.ps1 -useb | iex
# run project example:
deno run --unstable --allow-all example/main.ts
# run project example (web):
deno run --allow-all --unstable https://raw.githubusercontent.com/hviana/scriptNOP/master/example/main.ts
# bundle scriptNOP lib to any runtime or web browsers:
deno bundle mod.ts nop.js
# bundle scriptNOP lib to any runtime or web browsers (web):
deno bundle https://raw.githubusercontent.com/hviana/scriptNOP/master/mod.ts nop.js

Particularities of this implementation

The framework has several facilities for the developer. Conditions can be composed of SubConditions and can have parameters such as fuzzy parameters. Actions have a paradigm-independent nature and are represented directly by a reference to a procedure. This Action procedure can directly notify values to FBEs in order to make changes to the fact base, and notify and receive Notifications from Attributes in order to read the fact base.

To support all these features, the core of the original NOP was modified. Basically, the Instigations, Methods and Conditions entities were removed, and several new Notification paths were added. The Instigations and Methods were removed to make way for a paradigm-independent Action procedure. The Premises were removed because their concept was unified with the Conditions, and it is possible to parameterize a Condition to act as a Premise, being that this concept unification together with the composition of Conditions allow a more generic and powerful interface. For the composition between Conditions, Notifications from Conditions to Conditions were created. For the unified concept of Premises and Conditions, Notifications of Attributes for Conditions were created.

Actions, since they are a direct reference to a paradigm-independent procedure, need a drastic amount of modifications. When using imperative programming mechanisms, for example, it is unpredictable to know what the code execution flow will be since there may be, for example, code suspension mechanisms such as an IF, making it impossible to predict which Notifications will be sent and received by this Action procedure. To solve this problem, this Action procedure has access to routines that explicitly notify and receive Notifications. When needing an Attribute, there is an asynchronous routine named notifyAttrAndWaitReply that sends a Notification to the respective Attribute that symbolizes a request to its current value, and the Attribute sends to the Action procedure a Notification with its current value that is captured by the routine notifyAttrAndWaitReply. In the end, the notifyAttrAndWaitReply routine returns the current value of the Attribute inside the Action procedure. To notify FBEs, there is also a routine named notifyFBE, which transmits to an FBE a Notification containing a set of new values ​​for that FBE. It is also possible to send an explicit Notification to other Rules to activate them by a routine named notifyRule, also considering the Notifications of their respective Conditions.

The change of state of a notifying entity is always verified by its exit Notification value. In this way, Notifications from Rules to Actions, Actions to Attributes and from Attributes to Actions always use the "FORCE" mode, since the value of the Notification does not matter and only the fact that it exists represents a change of state

The Action procedure open a gap for possible temporal and structural redundancies in relation to the original NOP core, however, allows a very great ease for the programmer. It is possible, for example, to use native language resources such as the 'setTimeout' procedure to delay the execution of the Action procedure and implement a basic priority resource.

FBEs notify attributes in depth. That is, if an Attribute A2 is inside the Attribute A1 (i.e. A1.A2), Notifications about A2 will trigger Notifications about A1. If an Attribute is replaced by an entirely new Attribute (or created), its SubAttributes are also notified. It is possible to visualize the behavior in Code 1.

//initial values.
fbe.notify(
  {
    a: {
      b: {
        c: "foo",
      },
      d: true,
    },
  },
);
/*
Conditions/Premises that use "a", "a.b", "a.b.c", "e" and "e.f" will be notified.
Conditions/Premises that use only "a.d" are not notified, since
the value of "a.d" has not been modified.
*/
fbe.notify(
  {
    a: {
      b: {
        c: "bar",
      },
      d: true,
    },
    e: {
      f: 1,
    },
  },
);
fbe.get("a.b"); //returns "a.b" Attribute.

Code 1. Depth reactivity.

A better view of the inference flow driven by Notifications can be seen in Figure 2 compared to the original NOP core in Figure 1.

Figure 1. Flow of Notifications in the original NOP core.

Figure 2. Flow of Notifications in the scriptNOP core.

The framework also has a debugger procedure that intercepts all Notifications between NOP core entities. In this way, it is possible, for example, to save this information in a history or print it on the screen, in addition to allowing the implementation of explainability mechanisms.

Multithreaded and concurrent approach used

The multithreaded approach used has an asynchronous nature that eliminates the need for synchronization mechanisms between threads and control mechanisms such as mutexes, reducing the CPU overhead in parallelization. However, the approach cannot evenly distribute the load and tends to overload the main thread.

In this approach, there is a main thread named "state manager", which contains the entire base of facts (FBEs and Attributes) together with a global map containing the Notification between all entities. There are also N secondary threads, where each of these secondary threads contains the same set of all other entities like Conditions, Rules and Actions. The main thread, when receiving a Notification from any entity, checks if it causes a Notification map change given the value of the Notification, and if so, sends to one of the secondary threads an evaluation request for the recipient of this Notification, sending together all the other Notifications that were intended for that recipient (with the exception of FBEs and Attributes that have their state evaluated in the main thread). At the end of the evaluation, the secondary thread sends the exit Notification of the evaluated entity to the main thread, repeating the cycle.

The concept of concurrent programming is also explored. Operations in Conditions and Action procedures use asynchronous functions performed concurrently. In this way, if they use normally blocking operations such as I/O operations and HTTP requests, the thread is not blocked and is not idle.

The fact that the main thread intercepts all Notifications facilitates the implementation of mechanisms that use these Notifications to create debugger, history and explainability mechanisms.

Comparison of existing concepts in the literature and new ones in relation to materializations of the NOP

Many of the concepts used in this materialization of NOP are not new and already existed in old materializations. In Table 1 it is possible to visualize a list of these concepts and which ones are implemented [4].

Table 1. NOP concepts applied in materializations [4]. The concepts in bold were introduced in this materialization.

Data typing

The code is very dense, although every detail has been thought of in order to favor readability and avoid replication. With TypeScript, we have a new way of defining types and programming in an object-oriented style compared to classic object-oriented languages ​​such as Java and C++, which drastically reduces the amount of code. See the following code snippet:

export interface Attribute {
  fbe: string;
  attr: string;
}
// ...
export interface ConditionWithAttribute {
  attribute: Attribute;
}
export interface ConditionAsPremise {
  left: Condition;
  is: string;
  right: Condition;
}
// ...
export interface ConditionWithXor {
  xor: [Condition, Condition, ...Condition[]]; //min 2 Conditions
}
export type Condition =
  & (
    | ConditionWithAttribute
    | ConditionAsPremise
    | ConditionWithConstant
    | ConditionWithNot
    | ConditionWithAnd
    | ConditionWithOr
    | ConditionWithXor
    | ConditionWithFunc
    | ConditionWithRef
  )
  & FuzzyParameters;
//...
export class Rule {
  static coreLoaded: boolean = false;
  static initialized: boolean = false;
  #conditionTranspiler: ConditionTranspiler;

References

[1] J. M. Simão, C. A. Tacla, P. C. Stadzisz and R. F. Banaszewski, "Notification Oriented Paradigm (NOP) and Imperative Paradigm: A Comparative Study," Journal of Software Engineering and Applications, Vol. 5 No. 6, 2012, pp. 402-416. doi: https://www.doi.org/10.4236/jsea.2012.56047

[2] Melo, Luiz Carlos & Fabro, João & Simão, Jean. (2015). Adaptation of the Notification Oriented Paradigm (NOP) for the Development of Fuzzy Systems. Mathware& Soft Computing. 22. 1134-5632. url: https://www.researchgate.net/publication/279178301_Adaptation_of_the_Notification_Oriented_Paradigm_NOP_for_the_Development_of_Fuzzy_Systems

[3] F. Schütz, J. A. Fabro, C. R. E. Lima, A. F. Ronszcka, P. C. Stadzisz and J. M. Simão, "Training of an Artificial Neural Network with Backpropagation algorithm using Notification Oriented Paradigm," 2015 Latin America Congress on Computational Intelligence (LA-CCI), 2015, pp. 1-6, doi: https://doi.org/10.1109/LA-CCI.2015.7435978

[4] F. S. Neves, R. B. Linhares, "Framework NOP 4.0: contribution to the development of applications in the notification oriented paradigm through generic programming" (2021). Institutional Repository of the Federal Technological University of Paraná. url: https://repositorio.utfpr.edu.br/jspui/handle/1/26270

About

Author: Henrique Emanoel Viana, a Brazilian computer scientist, enthusiast of web technologies, cel: +55 (41) 99999-4664. URL: https://sites.google.com/site/henriqueemanoelviana

Improvements and suggestions are welcome!