Project 6: Processing data from a visual IoT sensor of fluid/structure interactions with machine learning REU Students: Pete Orkweha & Alexis Downing Graduate mentors: Nick Smith & Sharare Zehtabian Faculty mentor(s): Dr. Andrew Dickerson & Dr. Damla Turgut Week 4 (June 17 – June 21, 2019) Accomplishments: Data collection/analysis complete (240 random + 120 constant) Explored Poly-regressor, Random Forest Regressor (RFR), and Multi-layer Perceptron (MLP) on a complete dataset Literature research Problem & Solutions Problem: 𝑅 2 score doesn’t tell us everything. Need another way to score algorithms. Solution: We used RE as another way to score algorithms.
Optimizing Hyper-Parameters Random Search algorithm: Within a specify range, use random combination of hyper-parameters on an algorithm and return a combination that gives the best score. Grid Search algorithm Perform an exhaustive search for best combination of hyper-parameters with a given range of parameters. Results: MLP: Unoptimized score: 0.794 Optimized score: 0.905 RFR: no noticeable improvements
Varying Data set size Data size: 240 Starts at 3% (7 data) Increment by 3% K-fold = 5 folds
Machine learning algorithm result Poly-Regression: Each fold contains 47 data
Machine learning algorithm result MLP: Each fold contains 47 data
Machine learning algorithm result RFR: Each fold contains 47 data
Visualizing Algorithm Accuracy Poly-Regression
Max. Deflection vs Effective Length RFR algorithm Poly-Regression MLP algorithm
Max. Deflection vs Drop velocity RFR algorithm Poly-Regression MLP algorithm
Max. Deflection vs Drop Diameter RFR algorithm Poly-Regression MLP algorithm
Max. Deflection vs Moment of Inertia RFR algorithm Poly-Regression MLP algorithm
Project 6: Processing data from a visual IoT sensor of fluid/structure interactions with machine learning REU Students: Pete Orkweha & Alexis Downing Graduate mentors: Nick Smith & Sharare Zehtabian Faculty mentor(s): Dr. Andrew Dickerson & Dr. Damla Turgut Week 4 (June 17 – June 21, 2019) Plans for next week: Reducing features by combining features. Continue literature research Implement data pre-processing to improve algorithms Find out the relationship between drop velocity and deflection on fiber B3