Prudce inteligentní talk o umělé inteligenci

Talk pro PyVo, 30. června 2016

- úvod do AI
- počítání podle přírody aneb biocomputing

Transcript

3. Michal Wiglasz @kacer „Haven‘t used Python for serious business for about a year.“
5. Let‘s google a definition: An area of computer science that deals with giving machines the ability to seem like they have human intelligence. Simulation of human intelligence processes by machines.
7. Make the opponent suffer!
8. „We must go deeper.“ My move My move Opponents move These are final states with my score -1 0 1 Opponents move
9. These are final states with my score „There and back again“ (algorithms min-max, alfa-beta…) -1 1 1 1 1 -1 -1 My move towards MAX score My move towards MAX score Opponents move towards MIN score -1 0 1 Opponents move towards MIN score 1 0 0 0 0 -1 -1 1 1
11. My move DOWN „Opponent“ RIGHT
12. Like reading room numbers from door plates. https://github.com/kacer/dungeons-marathon 2012; openCV, pyBrain – still works, even on Windows
13. Blur, resize, lower resolution, crop, align, whatever floats your boat…
14. Found contours:
15. Found contours:
16. Interesting regions:
17. Resized to 40×40 pixels:
18. Classification: ML method which assigns classes to inputs Supervised learning Ok, computer, this is an „A“: Ok, computer, this is a „C“: Ok, computer, this is an „L“: Ok, human, I‘ll learn that! Input here: 40×40 (1600) binary values
19. Ok, computer, what is this? Hey, human, this is an „A“. Ok, computer, what is this? Hey, human, this is an „L“. Ok, computer, what is this? Hey, human, this is an „A“. Hey, human, not sure. „A“?
20. L327
21. L327 feedback feedback
22. 1) Import antigravity https://wiki.python.org/moin/PythonForArtificialIntelligence 2) Pick a toy problem 3) Dance like nobody‘s around!
23. http://playground.tensorflow.org/
32. Make babies Pick parents 32
33. 33 Make babies Pick parents
34. 1. Initialize the population 2. Pick the strongest ones 3. Make children, eliminate the weak ones 4. Repeat
35. Pick proper problem encoding! Chromosome (genotype) for graph coloring: Node 1 2 3 4 5 6 7 8 9 Color red red blue green red yellow blue green green
36. Fitness function: how good/bad is the individual In graph coloring: number of edges connecting nodes with the same color
37. Mom: Dad: Baby: Mutated baby Node 1 2 3 4 5 6 7 8 9 Color red red blue green red yellow blue green green Node 1 2 3 4 5 6 7 8 9 Color blue green red yellow yellow yellow red red green Node 1 2 3 4 5 6 7 8 9 Color red red blue green red yellow red red green Node 1 2 3 4 5 6 7 8 9 Color red blue blue green red yellow red red green
38. import random COLORS = ["red", "green", "blue", "yellow"] INDIVIDUALS = 10 NODES = 9 TARGET_FITNESS = 0 make_individual = lambda: [random.choice(COLORS) for _ in range(NODES)] population = [make_individual() for _ in range(INDIVIDUALS)] new_population = [None for _ in range(INDIVIDUALS)] generation = 0 while True: fitnesses = map(fitness_func, population) if max(fitnesses) == TARGET_FITNESS: break for i in range(INDIVIDUALS): mom, dad = pick_parents(population) split_point = random.randint(0, NODES - 1) new_population[i] = mutate(mom[:split_point] + dad[split_point:]) population = new_population generation += 1
39. All possible combinations Acceptable solutions Optimum Valid candidate solutions
40. All possible combinations Valid candidate solutions Acceptable solutions Optimum Accesible using conventional methods
41. All possible combinations Acceptable solutions Optimum Valid candidate solutions 1/6
42. All possible combinations Acceptable solutions Optimum Valid candidate solutions 2/6
43. All possible combinations Acceptable solutions Optimum Valid candidate solutions 3/6
44. All possible combinations Acceptable solutions Optimum Valid candidate solutions 4/6
45. All possible combinations Acceptable solutions Optimum Valid candidate solutions 5/6
46. All possible combinations Acceptable solutions Optimum Valid candidate solutions 6/6
47. All possible combinations Acceptable solutions Optimum Valid candidate solutions 1/5
48. All possible combinations Acceptable solutions Optimum Valid candidate solutions 2/5
49. All possible combinations Acceptable solutions Optimum Valid candidate solutions 3/5
50. All possible combinations Acceptable solutions Optimum Valid candidate solutions 4/5
51. All possible combinations Acceptable solutions Optimum Valid candidate solutions 5/5
52. All possible combinations Acceptable solutions Optimum Valid candidate solutions 1/4
53. All possible combinations Acceptable solutions Optimum Valid candidate solutions 2/4
54. All possible combinations Acceptable solutions Optimum Valid candidate solutions 3/4
55. All possible combinations Acceptable solutions Optimum Valid candidate solutions 4/4
56. Genetic algorithm Evolutionary strategies Evolutionary programming Genetic programming …
58. Velikost prediktoru Generace Trénovací obrázky 58
59. Velikost prediktoru Generace Trénovací obrázky 59
60. Velikost prediktoru Generace Trénovací obrázky 60
61. Velikost prediktoru Generace Trénovací obrázky 61
62. Velikost prediktoru Generace Trénovací obrázky 62
63. Velikost prediktoru Generace Trénovací obrázky 63
64. Velikost prediktoru Generace Trénovací obrázky 64
65. Velikost prediktoru Generace Trénovací obrázky 65
66. Velikost prediktoru Generace Trénovací obrázky 66
67. Velikost prediktoru Generace Trénovací obrázky 67
68. Velikost prediktoru Generace Trénovací obrázky 68
69. Velikost prediktoru Generace Trénovací obrázky 69
70. 10 % 30 % 70 % Salt&pepper noise 70
71. Edge detector 71