SP/Radio Primitive Recognition

journeycartAI and Robotics

Oct 15, 2013 (3 years and 9 months ago)

63 views

Semantic Signal Processing for Re
-
hosting
CR/SDR Implementations

SP/Radio Primitive Recognition


Jiadi Yu, Yingying Chen




1

SSP Framework

Abstract conceptual primitives
(“Thing, Place, Path, Action,
Cause”) from existing
implementations of signal
processing modules/systems
in source code

Represent the implementation
profile of signal processing
modules/systems based on
cognitive linguistics

Parse cognitive
-
linguistics
-
based representation and
generate implementation code
in the target platform

2

Radio
-
Level Abstraction


Abstract primitives at Radio
-
level


Analyze the Code
-
level primitives to recognize Radio
-
level
primitives

Algebraic

calculation:


+,
-
, *, /

Logic calculation:


xor, nor, and

Type conversions

Relational Operator


==
,!
=

Conditional control:


if… else…, while


:



Code level

Signal Sources

Signal Sinks

Filters

Signal Modulation

Signal Demodulation

Source coding

Synchronization

Equalization

AGC

OFDM locks



:



Radio level

Primitives of Semantic Radio

3


Radio
-
Level Abstraction (cont’)





Sources

Code

Radio level

XML

Presentation

Code level

XML

Presentation

Inference

Engine

Knowledge
Base

Radio

Primitives

Radio Level

Abstraction

Target

Code

Code level


Abstraction

SP module
recognition

4

Learning Based Inference Engine


Inference engine is able to understand the what level
primitives in the semantic presentation need to parsing



Inference engine is able to know what primitives need to
generate target code and what primitives just use code
from code library



Machine knows how to implement any
-
level primitives in
the target code


5

Learning Based Inference Engine

Inference Engine







Radio/Code

Presentation


Target

Code

Parser

Higher
-
level

Reinforcement
learning

Knowledge
Base

Learning

Agent

Information

Inquiry

Code

Generate

Conceptual
Primitives

lower
-
level

SP module
recognition

6

SP/Radio Primitive Recognition



Objective



Automated recognition of functionality of a
SP/Radio primitive


Automated recognition of functions from
knowledge library to perform desired action



Recognize the equivalence of two different
implementations


7

Primitive Recognition

-

Potential Approaches


Context
-
based


Function names


Comments



Behavior pattern


Tree
-
based pattern recognition


Machine learning
-
based pattern recognition



8

Context
-
based Recognition


Information retrieval


from
Function names/Comments


Function names


Direct comparison


Fuzzy matching and identification


Comments


Keyword
-
based


Machine learning models


9


The representation architecture based on cognitive linguistics
of the signal processing implementation is a
Tree Structure
.

Tree
-
based Pattern Recognition


Each signal processing
module can be
represented as a
behavior pattern using
lower
-
level primitives



Each signal processing
module can be
represented as a tree
architecture.






10


Tree
-
based Pattern Recognition


Primitive
Recognition







Tree
architecture
analyze

Knowledge
base

Tree
representation

Source

Target

11

An Example of QPSK



two QPSK implementations

Tree

representation

Binary Tree

representation

12

Tree
-
based Pattern Recognition

(Cont’)


Advantage



Direct comparison
Accuracy can be high



Disadvantage


C
ompare with all modules/functions of Knowledge
base
Slow, high computational cost



13

Machine Learning
-
based

Pattern Recognition



Based on the correlation between the radio primitive and identified
features



Potential Features


Lower
-
level primitives


Example: lookup table



Hierarchical architecture

-
Example: QPSK
includes a
lookup table primitive



Numerical attributes

-
Example:
integers, real numbers


Input/output variable types and ranges

-
Example: Input/output parameters of a filter is
array







14

A Simple Filter Example


The basic element

for the simple

filter include:

LOOP

ACCUMLATION


MULTIPLY

ARRAY

void main(){

for(i = 0; i < N ; i = i + 1){


k = N
-

i;


temp = tap[i] * input[k];


sum = sum + temp;


}

}

The code segments probably
implement functionality of


a filter

15

Machine Learning
-
based

Pattern Recognition

(Cont’)


Advantage



Fast & simple



Disadvantage


Accuracy can be low


16

ML and Tree
-
based Pattern
Recognition


Low computational cost and high accuracy

ML
-
based

Pattern Recognition

Tree
-
based

Pattern Recognition

First
step

Second
step

similar
primitives

Primitive Recognition






Source

Target

17

Thank You

Comments & Questions?

18