5. dotAI’18 - Virginie Mathivet – R&D Officer
First results
• Accuracy: 99.1%!!! (in 2 epochs)
6. dotAI’18 - Virginie Mathivet – R&D Officer
First results
• Accuracy: 99.1%!!! (in 2 epochs)
Real Predicted OK (1665) KO (15) Accuracy
OK 1665 0 100%
KO 15 0 0%
Total 1680 0 99.1%
• Confusion matrix:
15. dotAI’18 - Virginie Mathivet – R&D Officer
The choice of the evaluation function
Classification
Real Predicted 1 0
1 A (true positive) C (false negative)
0 B (false positive) D (true negative)
Accuracy Precision Recall / Sensitivity F1-score
(A+D)/(A+B+C+D) A/(A+B) A/(A+C) 2*precision*recall/(precision+recall)
Specificity Detection rate Youden’s J Index Balanced accuracy
D/(B+D) A/(A+B+C+D) Sensitivity+specificity - 1 (sensitivity+specificity) / 2
16. dotAI’18 - Virginie Mathivet – R&D Officer
The choice of the evaluation function
Regression
• R-squared (R2)
• Root Mean Squared Error (RMSE)
• Residual Standard Error (RSE) or
Model Sigma
• Mean Absolute Error (MAE)
• Akaike’s Information Criteria (AIC)
• Bayesian Information Criteria (BIC)
• …
17. dotAI’18 - Virginie Mathivet – R&D Officer
The choice of the evaluation function
0
5
10
15
20
25
0 1 2 3 4 5 6 7 8 9 10
Behavioral task
18. dotAI’18 - Virginie Mathivet – R&D Officer
Teach a (virtual) spider how to walk
• 1st try: distance made in 10sec
19. dotAI’18 - Virginie Mathivet – R&D Officer
Teach a (virtual) spider how to walk
• 1st try: distance made in 10sec
https://gymnasticselectronicportfolio13.weebly.com/gymnastics-skills.html
20. dotAI’18 - Virginie Mathivet – R&D Officer
Teach a (virtual) spider how to walk
• 1st try: distance made in 10sec
https://gymnasticselectronicportfolio13.weebly.com/gymnastics-skills.html
https://viatcheslav.livejournal.com/231564.html
21. dotAI’18 - Virginie Mathivet – R&D Officer
Teach a (virtual) spider how to walk
• 2nd try: distance made in 10sec without rolling
22. dotAI’18 - Virginie Mathivet – R&D Officer
Teach a (virtual) spider how to walk
• 2nd try: distance made in 10sec without rolling
https://blog.myagilepartner.fr/wp-content/uploads/2017/08/scared_bug.jpg
23. dotAI’18 - Virginie Mathivet – R&D Officer
Teach a (virtual) spider how to walk
• 2nd try: distance made in 10sec without rolling
https://www.memecenter.com/fun/1717889/flying-spiderhttps://blog.myagilepartner.fr/wp-content/uploads/2017/08/scared_bug.jpg
24. dotAI’18 - Virginie Mathivet – R&D Officer
Teach a (virtual) spider how to walk
• 3rd try: idem but after the bug correction
25. dotAI’18 - Virginie Mathivet – R&D Officer
Teach a (virtual) spider how to walk
• 3rd try: idem but after the bug correction
https://www.nemusloren.it/2013/11/30/come-se-fosse-possibile/
26. dotAI’18 - Virginie Mathivet – R&D Officer
Teach a (virtual) spider how to walk
• 3rd try: idem but after the bug correction
https://www.nemusloren.it/2013/11/30/come-se-fosse-possibile/
27. dotAI’18 - Virginie Mathivet – R&D Officer
Rule of thumb: speak to a child
Eat with a fork:
28. dotAI’18 - Virginie Mathivet – R&D Officer
Rule of thumb: speak to a child
Eat with a fork:
29. dotAI’18 - Virginie Mathivet – R&D Officer
Rule of thumb: speak to a child
Eat with a fork:
30. dotAI’18 - Virginie Mathivet – R&D Officer
Rule of thumb: speak to a child
Eat with a fork:
33. dotAI’18 - Virginie Mathivet – R&D Officer
Overfitting spotting
Error
Time
Training
Validation
Underfitting Overfitting
Optimum
34. dotAI’18 - Virginie Mathivet – R&D Officer
Teaching a robot how to avoid walls
• Evaluation function :
• 1st try: time until a collision with a wall (max 10sec)
35. dotAI’18 - Virginie Mathivet – R&D Officer
Teaching a robot how to avoid walls
• Evaluation function :
• 1st try: time until a collision with a wall (max 10sec)
36. dotAI’18 - Virginie Mathivet – R&D Officer
Teaching a robot how to avoid walls
• Evaluation function :
• 1st try: time until a collision with a wall (max 10sec)
• 2nd try: time moving until a collision with a wall (max 10sec)
37. dotAI’18 - Virginie Mathivet – R&D Officer
Teaching a robot how to avoid walls
• Evaluation function :
• 1st try: time until a collision with a wall (max 10sec)
• 2nd try: time moving until a collision with a wall (max 10sec)
38. dotAI’18 - Virginie Mathivet – R&D Officer
Teaching a robot how to avoid walls
• Evaluation function :
• 1st try: time until a collision with a wall (max 10sec)
• 2nd try: time moving until a collision with a wall (max 10sec)
• 3rd try: time moving forward until a collision with a wall (max 10sec)
39. dotAI’18 - Virginie Mathivet – R&D Officer
Teaching a robot how to avoid walls
• Evaluation function :
• 1st try: time until a collision with a wall (max 10sec)
• 2nd try: time moving until a collision with a wall (max 10sec)
• 3rd try: time moving forward until a collision with a wall (max 10sec)
40. dotAI’18 - Virginie Mathivet – R&D Officer
Theory vs Reality
• Real robot and real cage => not working at all!!
• Unseen overfitting!! (no dataset = no curves)
41. dotAI’18 - Virginie Mathivet – R&D Officer
Theory vs Reality
• Real robot and real cage => not working at all!!
• Unseen overfitting!! (no dataset = no curves)
42. dotAI’18 - Virginie Mathivet – R&D Officer
Theory vs Reality
• Real robot and real cage => not working at all!!
• Unseen overfitting!! (no dataset = no curves)
• Solution :
• Randomness
• Test on real robots as soon as possible