AI-FML for Inference of Percentage of Votes Obtained and Real-life - PowerPoint PPT Presentation

AI-FML for Inference of Percentage of Votes Obtained and Real-life Applications Team : Yangming and Dominic Presenters: Ying-Tai Chen, Seng-Yo Huang From: St. Dominic Catholic High School Yangming junior high school , Kaohsiung, Taiwan

Outline  Introduction  Machine learning Flowchart  FML-based Knowledge Base & Rule Base  Fuzzy Inference Mechanism  Machine Learning – Genetic and PSO Algorithm  Testing  Conclusions

Introduction  By using the FML intellectual decision tool to establish the knowledge base and rule base of the fuzzy inference system; optimize the FML knowledge base and rule base through the methods of machine learning.  Machine learning is executed by means of regression, ensemble methods, and deep learning.

Goal  The goal of our project is to create a model used to predict the final outcome of the election, an make those of the predicted results closer to the actual outcome of the election.  To compare the difference of the two algorithms, GA and PSO, and perform testing based on the results of the elections.  Perform hard/software integration and utilize robotic applications

Machine Learning Process Feature extraction Original data Training data Normalization A. Prediction model Machine Testing data learning B. Decision model

FML Fuzzy Markup Language

Establishment of FML-Based Knowledgebase  Input  PA(Power Area)  AD(Age Distribution)  PS(Poll Satisfaction)  PE(Political Experience)  Output  ET(Election)

PA(Power Area)  Denotes the interdependence of the candidate’s strongholds and the electorate since it enhances emotional identification  A bonus points system is used  Three points is given when the candidate is born in the electorate  Two points is given when the candidate is born in a different place, but the candidate’s spouse is born in the same place, and  One point is given when there is no direct relativity between the candidate and the electorate

AD(Age Distribution)  According to findings of recent elections, young votes dramatically influence the final outcome; whereas older people would have a rather fixed political preference  We used weighted mean to convert data to AD  20-29 yrs. Old multiplied by 5  30-39yrs. Old multiplied by 4  40-49yrs. Old multiplied by 3  50-59yrs. Old multiplied by 3  Beyond 60yrs. Old multiplied by 3

PS(Poll Satisfaction)  The Credits/ Discredits will directly affect the votes received by the candidate  PS is calculated according to the average of the last conducted poll survey

PE(Political Experience)  Denotes the political experience of the candidate since the constituencies’ decisions are often based on the similar political experiences of the candidate  A bonus points system is used  Plus one point if worked as chief of village  Plus two points if worked as member of parliament or Members of the Legislative Yuan  Plus three points if worked as commissioner  Plus four points if worked as deputy mayor  Plus five points if worked as mayor of the city

ET(Election)  The actual votes received by the candidate

KB and RB — KB Criteria (Before Learning)  PA 、 AD  PE  Low:[0,0.1,0.3,0.4]  Low:[0,0.2,0.4,0.6]  Medium:[0.3,0.4,0.6,0.7]  Medium_Low:[0.4,0.6,0.8,1]  High:[0.6,0.7,0.9,1]  Medium_High:[0.8,1,1.4,1.6]  High:[1.4,1.6,1.8,2]  PS  ET  Low:[0,0.1,0.2,0.3]  Medium_Low:[0.2,0.3,.4,0.5]  Low:[0,0.1,0.2,0.3]  Medium_High:[0.4,0.5,0.7,0.8]  Medium_Low:[0.2,0.3,0.4,0.5]  High:[0.7,0.8,0.9,1]  Medium:[0.4,0.5,0.6,0.7]  Medium_High:[0.6,0.7,0.8,0.9]  High:[0.8,0.9,1,1]

Principle of Knowledge Base

Figure of Knowledge Base(Before Learning)

Rule Base

Rule Base Diagram

Selection and Normalization  Selection: We gathered data from numerous elections across Taiwan, and 66 of them became training data, and the rest of them became testing data(22). Normalization: PA/PE: Divide the whole column by 10. AD/PS/ET: Divide the whole column by 100 .

Selection and Normalization Orig. Aft.

Training Data and Testing Data Output Input

FML Machine Learning — GA  Using GA(Genetic Algorithm)  Establish the knowledge base and input the rule base  Using GA to perform machine learning(Training Data) and receive an optimized model  Perform testing(Testing Data)

FML learning interface

Genetic Learning Results Generation

Figure of KB (After GA Learning)

Training Results

FML Inference Interface

Correlation Coefficient  The correlation coefficient is a statistical measure of the strength of the relationship between the relative movements of two variables. The values range between -1.0 and 1.0. A calculated number greater than 1.0 or less than - 1.0 means that there was an error in the correlation measurement. A correlation of -1.0 shows a perfect negative correlation, while a correlation of 1.0 shows a perfect positive correlation. A correlation of 0.0 shows no linear relationship between the movement of the two variables.

Inference Results--By 10000 Generation Model Inference Data 0.9 0.8 0.7 0.6 0.5 correl 0.842571 0.4 0.3 0.2 0.1 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65 ET ET aft.

Inference Results--By 7000 Generation Model Inference Data 0.8 correl 0.898741 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65 ET ET aft.

Testing Results — Internal Test internal test 0.8 correl 0.850459 0.7 0.6 0.5 0.4 0.3 0.2 0.1 0 1 3 5 7 9 11 13 15 17 19 21 23 25 27 29 31 33 35 37 39 41 43 45 47 49 51 53 55 57 59 61 63 65 ET ET orig.

Testing Results — External Test No PA AD PS PE ET ET orig. 1 0 0.130556 0.152 0.4 0 0.4586 2 0.3 0.475556 0.58 0.5 0.409 0.5414 3 0.3 0.057778 0.05 0.4 0 0.1095 external test 4 0.3 0.510556 0.47 1 0.523 0.4948 5 0.3 0.172222 0.19 0.6 0.255 0.3823 0.7 6 0.3 0.237222 0.27 0.4 0.274 0.382 7 0.3 0.155 0.09 0 0 0.2768 0.6 8 0.3 0.292778 0.31 0.4 0.274 0.3229 9 0.3 0.161667 0.15 0.4 0.272 0.1798 0.5 10 0.3 0.356111 0.4 1 0.348 0.4118 11 0.1 0.02 0.01 0 0 0.0128 0.4 12 0.3 0.292222 0.41 0.2 0.274 0.3956 13 0.1 0.049444 0.046 0.5 0 0.0255 0.3 14 0.3 0.343333 0.31 0.8 0.266 0.3704 15 0.3 0.005 0.013 0 0 0.0054 0.2 16 0.3 0.001 0.04 0 0 0.0083 17 0.3 0.411667 0.4 0.6 0.517 0.5905 0.1 18 0.1 0.606111 0.59 0.6 0.515 0.5653 19 0.1 0.158889 0.17 0.2 0.274 0.2762 0 20 0.1 0.095556 0.07 0 0 0.0549 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 21 0.3 0.462222 0.44 1.1 0.527 0.6309 22 0.3 0.145556 0.15 0.8 0.245 0.3409 ET ET orig. correl 0.846211

2020 Kaohsiung Mayor Election Inference Data 2020 Election Testing 0.8 ET orig. ET aft. 0.7 0.7003 0.7 0.6 0.0406 0 0.5 0.4 0.259 0.259 0.3 0.2 0.1 correl 0.99904716 0 1 2 3 ET orig. ET aft.

FML Machine Learning — PSO  Using PSO(Particle swarm optimization)  Establish the knowledge base and input the rule base  Using PSO to perform machine learning(Training Data) and receive an optimized model  Perform testing(Testing Data)

KB Diagram

RB Diagram

PSO Algorithm for Testing Data(10G)

PSO Algorithm for Testing Data(100G)

KB Diagram(After learning)

Fuzzy Inference(PSO)

Fitness/Accuracy

2020 Kaohsiung Mayor Election Inference Data

2020 Kaohsiung Mayor Election Inference Data PSO Algorithm 0.8 0.7 ET orig. ET AFT . 0.6 0.7003 0.50059873 0.5 0.0406 0.09172902 0.4 0.259 0.32717 0.3 0.2 0.1 CORREL 0.96105583 0 1 2 3 ET orig. ET AFT

Connection with Kebbi robot — Blockly Input the variables for Kebbi

The Kebbi AI bot will vocally tell us the output

Comparison Between GA and PSO  Compare GA and PSO  A curvature design is shown in the KB diagram of GA, but PSO doesn’t have this function up to now. And GA is also better since it shows a fuzzy concept during learning and inference  PSO has a colored design in the KB diagram which is friendlier to new starters  After Fuzzy inference or fitness/accuracy calculations, PSO directly indicates the corresponding level, therefore it’s easier to use speaking of this function.

Conclusion  According to the Internet, we could get reliable data and therefore predict the final results of the election  After comparing GA and PSO algorithm, we found that PSO is more suitable for new starters; however, speaking of performance and accuracy, GA is better.  Correlation is found within the two lists of testing results which is further quantified by the correlation coefficient  The isn’t always a relation between accuracy and generations