際際滷

際際滷Share a Scribd company logo

Machine Learning for Gamers - Dungeon Forecasts & Dragon Regressions

Guy Royse
Guy Royse

Its Friday night and youre making your character for a fun evening of gaming. Youve rolled your abilities and even got two 18s. But what class should you pick given your rolls? If you pick that class, is your character playable? What kind of character have you created? Or maybe youre a game master and youve been running a game for some time. Do you need to know how much treasure that dragon horde should have? Or are you wanting to figure out how many encounters your players will make it through in an evening so you can prepare enough material? These important questions can all be answered using machine learning. Many developers want to make use of machine learning in their applications but arent sure what sorts of problems can be solved with it. This talk will explain the sorts of problems that can be solved, what data is required to solve them, and what the results look like. And, well explore it using fun and geeky examples. We will cover six major types of problems that machine learning can solve: regression, forecasting, impact analysis, classification, clustering, and anomaly detection. When were done, youll have a basic understanding of what machine learning can do and what you might want to use it for. I might even be something other than role-playing!

1 of 44
Download to read offline
Machine Learning for Gamers Dungeon Forecasts and Dragon Regressions
Who is this Guy? Guy Royse Developer Evangelist Nexosis guy@nexosis.com @guyroyse 2
Who is this Company? 3 Nexosis is a company of developers focused on providing easy access to machine learning.
4 IANADS
5
6 What is Machine Learning?
7 Gather Data Build a Model Make Predictions
8 Gather Data Target Features Encoding Imputation Build a Model Make Predictions
Target & Features Target The thing we want to predict Features Things that a鍖ect the thing we want to predict 9
Encoding & Imputation 10 Encoding Converting non-numeric data to numeric data Imputation Replacing nulls in the data with meaningful replacements
11 Encoding Gobliniod Yes No Goblinoid 1 0 Color Black Blue Green Red White Black Blue Green Red White 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 Boolean Encoding One-Hot Encoding
12 Gather Data Build a Model Algorithm Train Test Make Predictions
Algorithms 13 ARIMA Various combinations of AutoRegressive component with p parameters Di鍖erencing component with d parameters Moving Average component with qparameters with external regressors Exponential Smoothing Simple Double Triple with Box-Cox Transformation Autoregressive Neural Network with or without external regressors Multiple Linear Regression with or without external regressors Spline Seasonal and Trend Decomposition using Loess with ARIMA with or without external regressors with Exponential Smoothing Bayesian Time Series Regression with or without external regressors Additive Model Home-grown Nexosis Algorithms Least Squares Linear Polynomial Elastic Net Lasso Ridge Support Vector Regression Linear Kernel Polynomial Kernel Radial Basis Function kernel Sigmoid Kernel Multi-Layer Perceptron (Neural Network) with 1, 2, or 3 hidden layers Recti鍖ed Linear Unit Function Hyperbolic Tan Function Sigmoid Function Random Forest K-Nearest Neighbor Logistic Regression Naive Bayes
14 Train & Test 467 366 832 915 27 499 217 693 659 506 411 645 677 50 514 251 198 792 398 418 781 297 77 124 406 967 728 295 344 483 332 983 734 949 270 891 509 426 880 163 67 407 63 227 713 152 2 247 956 530 407 442 103 558 36 762 727 79 172 50 810 389 374 607 215 386 7 927 906 119 743 480 577 226 160 395 635 745 117 391 753 570 106 366 470 228 771 524 973 136 844 334 922 332 297 129 17 154 41 198 862 773 165 679 241 742 731 603 366 866 872 20 891 742 933 578 59 723 758 785 153 336 84 829 851 210 144 140 747 306 214 121
15 Gather Data Build a Model Algorithm Train Test Make Predictions
16 Gather Data Build a Model Make Predictions Provide Features Get Target
17 Regression Classification Anomaly Detection Forecasting Impact Analysis
18 Regression Determine the value of a dragons hoard in gold pieces based on the age, color, and hit points of that dragon.
19 About Dragons Wyrmling Young Adult Ancient Black Blue Green Red White Brass Bronze Copper Gold Silver How much damage the dragon can take until defeated Total value of all the dragons treasure in gold pieces Age Color Hit Points Hoard Value
20 Training Data Age Color Hit Points Hoard Value Young Gold 178 3,419.31 Ancient Blue 481 105,630.42 Ancient Green 385 107,355.23 Wyrmling Green 38 233.15 Adult Red 256 152,685.62 Adult Brass 172 4,490.94 Young Silver 168 2,786.95 Wyrmling Copper 22 155.11 Young Black 127 5,345.34 Adult White 200 3,789.23 Wyrmling Bronze 32 556.12 Ancient Bronze 444 123,891.74 Adult White 223 10,345.45
21 Asking the Question Age Color Hit Points Hoard Value Ancient Gold 527 ? Age Color Hit Points Hoard Value Ancient Gold ? 129,459.14
22 Can I Ask This? Age Color Hit Points Hoard Value Ancient ? 527 129,459.14 Age Color Hit Points Hoard Value ? Gold 527 129,459.14
23 Classification Based on the six key abilities and race of a character, determine what class they should play.
24 About Characters Strength Dexterity Constitution Intelligence Wisdom Charisma Elf Dwarf Halfling Human Gnome Half-Elf Half-Orc Tiefling Barbarian Bard Cleric Druid Fighter Monk Paladin Ranger Rogue Sorcerer Warlock Wizard Abilities Race Class Dragonborn
25 Training Data Str Dex Con Int Wis Cha Race Class 14 13 15 11 9 15 Half-Orc Barbarian 8 12 15 13 11 12 Half-Elf Rogue 12 10 15 16 5 10 Elf Wizard 17 17 17 10 7 18 Dwarf Fighter 15 15 14 11 11 13 Human Fighter 13 8 14 5 12 17 Halfing Bard 18 16 15 13 9 15 Halfling Rogue 10 10 15 12 15 16 Tiefling Warlock 18 11 10 12 12 11 Dwarf Fighter 16 5 9 8 16 16 Human Paladin 16 14 11 13 13 15 Dragonborn Fighter 5 9 14 9 11 13 Human Bard 10 12 7 15 15 10 Gnome Cleric
26 Asking the Question Str Dex Con Int Wis Cha Race Class 14 13 15 11 9 15 Half-Orc ? Barbarian 0.97184 Bard 0.83836 Cleric 0.81324 Druid 0.76682 Fighter 0.76646 Monk 0.64012 Paladin 0.5957 Ranger 0.5273 Rogue 0.44096 Sorcerer 0.27116 Warlock 0.18702 Wizard 0.07035
27 Asking the Question Str Dex Con Int Wis Cha Race Class 14 13 15 11 9 15 ? Cleric Dwarf 0.97184 Elf 0.83836 Halfling 0.81324 Human 0.76682 Dragonborn 0.76646 Gnome 0.64012 Half-Elf 0.27116 Half-Orc 0.18702 Tiefling 0.07035
28 Anomaly Detection Based on the six key abilities, race, and class of a character, determine if that character is playable or not.
29 Training Data Str Dex Con Int Wis Cha Race Class 14 13 15 11 9 15 Half-Orc Barbarian 8 12 15 13 11 12 Half-Elf Rogue 12 10 15 16 5 10 Elf Wizard 17 17 17 10 7 18 Dwarf Fighter 3 15 14 11 11 13 Human Fighter 13 8 14 5 12 17 Halfing Bard 18 16 15 13 9 15 Halfling Rogue 10 10 15 12 15 16 Tiefling Warlock 18 11 10 12 12 11 Dwarf Fighter 16 5 9 8 16 16 Human Paladin 16 14 11 13 13 15 Dragonborn Fighter 5 9 14 9 11 3 Human Bard 10 12 7 15 15 10 Gnome Cleric
30 Training Data Str Dex Con Int Wis Cha Race Class Playability 14 13 15 11 9 15 Half-Orc Barbarian 0.18 8 12 15 13 11 12 Half-Elf Rogue 0.06 12 10 15 16 5 10 Elf Wizard 0 17 17 17 10 7 18 Dwarf Fighter 0.36 3 15 14 11 11 13 Human Fighter -0.02 13 8 14 5 12 17 Halfing Bard 0.02 18 16 15 13 9 15 Halfling Rogue 0.36 10 10 15 12 15 16 Tiefling Warlock 0.2 18 11 10 12 12 11 Dwarf Fighter 0.12 16 5 9 8 16 16 Human Paladin 0.04 16 14 11 13 13 15 Dragonborn Fighter 0.28 5 9 14 9 11 3 Human Bard -0.34 10 12 7 15 15 10 Gnome Cleric 0.02
Str Dex Con Int Wis Cha Race Class Playability 16 14 16 6 9 4 Half-Orc Barbarian ? 8 8 8 18 16 10 Half-Orc Barbarian ? 31 Asking the Question
32 Forecasting Predict the number of encounters that will be completed per weekly session for future weeks based on which players will be present and how long the session will last.
33 About Gaming Sessions When the session occurred The players who attended the game for that particular session Time spent playing during that particular game session Number of completed combats for the session Date Players Length Encounters
34 Training Data Date Alice Bob Chuck Eve Length Encounters 11/13/2016 Yes Yes Yes Yes 4.00 4 11/20/2016 Yes No Yes Yes 4.50 4 11/27/2016 Yes Yes Yes Yes 3.25 3 12/4/2016 Yes Yes Yes Yes 5.50 6 12/11/2016 Yes Yes Yes No 12.00 8 12/18/2016 Yes No No Yes 4.00 7 12/25/2016 No No No No 0.00 0 1/1/2017 No No Yes Yes 4.50 3 1/8/2017 Yes Yes Yes Yes 7.00 5 1/15/2017 Yes Yes Yes Yes 6.25 7 1/22/2017 Yes Yes No Yes 6.00 8 1/29/2017 Yes Yes Yes No 4.00 4 2/5/2017 Yes Yes Yes Yes 3.75 3
Date Alice Bob Chuck Eve Length Encounters 2/12/2017 Yes Yes Yes Yes ? 4 2/19/2017 Yes Yes Yes Yes ? 5 2/26/2017 Yes Yes Yes No ? 4 Date Alice Bob Chuck Eve Length Encounters 2/12/2017 Yes Yes Yes Yes 4.00 ? 2/19/2017 Yes Yes Yes Yes 4.00 ? 2/26/2017 Yes Yes Yes No 6.00 ? 35 Asking the Question
36 Impact Analysis Measure the impact of updating your campaign from Pathfinder to 5th Edition.
37 Training Data Date Alice Bob Chuck Eve Length Encounters 11/13/2016 Yes Yes Yes Yes 4.00 4 11/20/2016 Yes No Yes Yes 4.50 4 11/27/2016 Yes Yes Yes Yes 3.25 3 12/4/2016 Yes Yes Yes Yes 5.50 6 12/11/2016 Yes Yes Yes No 12.00 8 12/18/2016 Yes No No Yes 4.00 7 12/25/2016 No No No No 0.00 0 1/1/2017 No No Yes Yes 4.50 3 1/8/2017 Yes Yes Yes Yes 7.00 5 1/15/2017 Yes Yes Yes Yes 6.25 7 1/22/2017 Yes Yes No Yes 6.00 8 1/29/2017 Yes Yes Yes No 4.00 4 2/5/2017 Yes Yes Yes Yes 3.75 3
Date Alice Bob Chuck Eve Length Encounters 1/1/2017 No No Yes Yes 4.50 3 1/8/2017 Yes Yes Yes Yes 7.00 5 1/15/2017 Yes Yes Yes Yes 6.25 7 1/22/2017 Yes Yes No Yes 6.00 8 1/29/2017 Yes Yes Yes No 4.00 4 2/5/2017 Yes Yes Yes Yes 3.75 3 38 Asking the Question Date Alice Bob Chuck Eve Length Encounters 1/1/2017 No No Yes Yes 4.50 ? 1/8/2017 Yes Yes Yes Yes 7.00 ? 1/15/2017 Yes Yes Yes Yes 6.25 ? 1/22/2017 Yes Yes No Yes 6.00 ? 1/29/2017 Yes Yes Yes No 4.00 ? 2/5/2017 Yes Yes Yes Yes 3.75 ?
39 Asking the Question Date Actual Encoutners Predicted Encounters Change 1/1/2017 3 5.0346 2.0346 1/8/2017 5 8.391 3.391 1/15/2017 7 11.7474 4.7474 1/22/2017 8 13.4256 5.4256 1/29/2017 4 6.7128 2.7128 2/5/2017 3 5.0346 2.0346
40 Regression Classification Anomaly Detection Forecasting Impact Analysis
41 Questions?
Who is this Guy? Guy Royse Developer Evangelist Nexosis guy@nexosis.com @guyroyse 42
43 Credits o Images o https://www.flickr.com/photos/t17emma/5909350021 o https://www.flickr.com/photos/rueink/35095791331 o https://www.flickr.com/photos/rueink/37125691160 o https://www.flickr.com/photos/fauxlaroid/5988373053 o https://www.flickr.com/photos/kimonomania/602858091 o https://www.flickr.com/photos/wonderferret/174314534 o https://www.flickr.com/photos/herval/4100847043
Machine Learning for Gamers - Dungeon Forecasts & Dragon Regressions

More Related Content

Machine Learning for Gamers - Dungeon Forecasts & Dragon Regressions

  • 1. Machine Learning for Gamers Dungeon Forecasts and Dragon Regressions
  • 2. Who is this Guy? Guy Royse Developer Evangelist Nexosis guy@nexosis.com @guyroyse 2
  • 3. Who is this Company? 3 Nexosis is a company of developers focused on providing easy access to machine learning.
  • 5. 5
  • 6. 6 What is Machine Learning?
  • 9. Target & Features Target The thing we want to predict Features Things that a鍖ect the thing we want to predict 9
  • 10. Encoding & Imputation 10 Encoding Converting non-numeric data to numeric data Imputation Replacing nulls in the data with meaningful replacements
  • 11. 11 Encoding Gobliniod Yes No Goblinoid 1 0 Color Black Blue Green Red White Black Blue Green Red White 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 0 0 0 0 0 1 Boolean Encoding One-Hot Encoding
  • 13. Algorithms 13 ARIMA Various combinations of AutoRegressive component with p parameters Di鍖erencing component with d parameters Moving Average component with qparameters with external regressors Exponential Smoothing Simple Double Triple with Box-Cox Transformation Autoregressive Neural Network with or without external regressors Multiple Linear Regression with or without external regressors Spline Seasonal and Trend Decomposition using Loess with ARIMA with or without external regressors with Exponential Smoothing Bayesian Time Series Regression with or without external regressors Additive Model Home-grown Nexosis Algorithms Least Squares Linear Polynomial Elastic Net Lasso Ridge Support Vector Regression Linear Kernel Polynomial Kernel Radial Basis Function kernel Sigmoid Kernel Multi-Layer Perceptron (Neural Network) with 1, 2, or 3 hidden layers Recti鍖ed Linear Unit Function Hyperbolic Tan Function Sigmoid Function Random Forest K-Nearest Neighbor Logistic Regression Naive Bayes
  • 14. 14 Train & Test 467 366 832 915 27 499 217 693 659 506 411 645 677 50 514 251 198 792 398 418 781 297 77 124 406 967 728 295 344 483 332 983 734 949 270 891 509 426 880 163 67 407 63 227 713 152 2 247 956 530 407 442 103 558 36 762 727 79 172 50 810 389 374 607 215 386 7 927 906 119 743 480 577 226 160 395 635 745 117 391 753 570 106 366 470 228 771 524 973 136 844 334 922 332 297 129 17 154 41 198 862 773 165 679 241 742 731 603 366 866 872 20 891 742 933 578 59 723 758 785 153 336 84 829 851 210 144 140 747 306 214 121
  • 18. 18 Regression Determine the value of a dragons hoard in gold pieces based on the age, color, and hit points of that dragon.
  • 19. 19 About Dragons Wyrmling Young Adult Ancient Black Blue Green Red White Brass Bronze Copper Gold Silver How much damage the dragon can take until defeated Total value of all the dragons treasure in gold pieces Age Color Hit Points Hoard Value
  • 20. 20 Training Data Age Color Hit Points Hoard Value Young Gold 178 3,419.31 Ancient Blue 481 105,630.42 Ancient Green 385 107,355.23 Wyrmling Green 38 233.15 Adult Red 256 152,685.62 Adult Brass 172 4,490.94 Young Silver 168 2,786.95 Wyrmling Copper 22 155.11 Young Black 127 5,345.34 Adult White 200 3,789.23 Wyrmling Bronze 32 556.12 Ancient Bronze 444 123,891.74 Adult White 223 10,345.45
  • 21. 21 Asking the Question Age Color Hit Points Hoard Value Ancient Gold 527 ? Age Color Hit Points Hoard Value Ancient Gold ? 129,459.14
  • 22. 22 Can I Ask This? Age Color Hit Points Hoard Value Ancient ? 527 129,459.14 Age Color Hit Points Hoard Value ? Gold 527 129,459.14
  • 23. 23 Classification Based on the six key abilities and race of a character, determine what class they should play.
  • 25. 25 Training Data Str Dex Con Int Wis Cha Race Class 14 13 15 11 9 15 Half-Orc Barbarian 8 12 15 13 11 12 Half-Elf Rogue 12 10 15 16 5 10 Elf Wizard 17 17 17 10 7 18 Dwarf Fighter 15 15 14 11 11 13 Human Fighter 13 8 14 5 12 17 Halfing Bard 18 16 15 13 9 15 Halfling Rogue 10 10 15 12 15 16 Tiefling Warlock 18 11 10 12 12 11 Dwarf Fighter 16 5 9 8 16 16 Human Paladin 16 14 11 13 13 15 Dragonborn Fighter 5 9 14 9 11 13 Human Bard 10 12 7 15 15 10 Gnome Cleric
  • 26. 26 Asking the Question Str Dex Con Int Wis Cha Race Class 14 13 15 11 9 15 Half-Orc ? Barbarian 0.97184 Bard 0.83836 Cleric 0.81324 Druid 0.76682 Fighter 0.76646 Monk 0.64012 Paladin 0.5957 Ranger 0.5273 Rogue 0.44096 Sorcerer 0.27116 Warlock 0.18702 Wizard 0.07035
  • 27. 27 Asking the Question Str Dex Con Int Wis Cha Race Class 14 13 15 11 9 15 ? Cleric Dwarf 0.97184 Elf 0.83836 Halfling 0.81324 Human 0.76682 Dragonborn 0.76646 Gnome 0.64012 Half-Elf 0.27116 Half-Orc 0.18702 Tiefling 0.07035
  • 28. 28 Anomaly Detection Based on the six key abilities, race, and class of a character, determine if that character is playable or not.
  • 29. 29 Training Data Str Dex Con Int Wis Cha Race Class 14 13 15 11 9 15 Half-Orc Barbarian 8 12 15 13 11 12 Half-Elf Rogue 12 10 15 16 5 10 Elf Wizard 17 17 17 10 7 18 Dwarf Fighter 3 15 14 11 11 13 Human Fighter 13 8 14 5 12 17 Halfing Bard 18 16 15 13 9 15 Halfling Rogue 10 10 15 12 15 16 Tiefling Warlock 18 11 10 12 12 11 Dwarf Fighter 16 5 9 8 16 16 Human Paladin 16 14 11 13 13 15 Dragonborn Fighter 5 9 14 9 11 3 Human Bard 10 12 7 15 15 10 Gnome Cleric
  • 30. 30 Training Data Str Dex Con Int Wis Cha Race Class Playability 14 13 15 11 9 15 Half-Orc Barbarian 0.18 8 12 15 13 11 12 Half-Elf Rogue 0.06 12 10 15 16 5 10 Elf Wizard 0 17 17 17 10 7 18 Dwarf Fighter 0.36 3 15 14 11 11 13 Human Fighter -0.02 13 8 14 5 12 17 Halfing Bard 0.02 18 16 15 13 9 15 Halfling Rogue 0.36 10 10 15 12 15 16 Tiefling Warlock 0.2 18 11 10 12 12 11 Dwarf Fighter 0.12 16 5 9 8 16 16 Human Paladin 0.04 16 14 11 13 13 15 Dragonborn Fighter 0.28 5 9 14 9 11 3 Human Bard -0.34 10 12 7 15 15 10 Gnome Cleric 0.02
  • 31. Str Dex Con Int Wis Cha Race Class Playability 16 14 16 6 9 4 Half-Orc Barbarian ? 8 8 8 18 16 10 Half-Orc Barbarian ? 31 Asking the Question
  • 32. 32 Forecasting Predict the number of encounters that will be completed per weekly session for future weeks based on which players will be present and how long the session will last.
  • 33. 33 About Gaming Sessions When the session occurred The players who attended the game for that particular session Time spent playing during that particular game session Number of completed combats for the session Date Players Length Encounters
  • 34. 34 Training Data Date Alice Bob Chuck Eve Length Encounters 11/13/2016 Yes Yes Yes Yes 4.00 4 11/20/2016 Yes No Yes Yes 4.50 4 11/27/2016 Yes Yes Yes Yes 3.25 3 12/4/2016 Yes Yes Yes Yes 5.50 6 12/11/2016 Yes Yes Yes No 12.00 8 12/18/2016 Yes No No Yes 4.00 7 12/25/2016 No No No No 0.00 0 1/1/2017 No No Yes Yes 4.50 3 1/8/2017 Yes Yes Yes Yes 7.00 5 1/15/2017 Yes Yes Yes Yes 6.25 7 1/22/2017 Yes Yes No Yes 6.00 8 1/29/2017 Yes Yes Yes No 4.00 4 2/5/2017 Yes Yes Yes Yes 3.75 3
  • 35. Date Alice Bob Chuck Eve Length Encounters 2/12/2017 Yes Yes Yes Yes ? 4 2/19/2017 Yes Yes Yes Yes ? 5 2/26/2017 Yes Yes Yes No ? 4 Date Alice Bob Chuck Eve Length Encounters 2/12/2017 Yes Yes Yes Yes 4.00 ? 2/19/2017 Yes Yes Yes Yes 4.00 ? 2/26/2017 Yes Yes Yes No 6.00 ? 35 Asking the Question
  • 36. 36 Impact Analysis Measure the impact of updating your campaign from Pathfinder to 5th Edition.
  • 37. 37 Training Data Date Alice Bob Chuck Eve Length Encounters 11/13/2016 Yes Yes Yes Yes 4.00 4 11/20/2016 Yes No Yes Yes 4.50 4 11/27/2016 Yes Yes Yes Yes 3.25 3 12/4/2016 Yes Yes Yes Yes 5.50 6 12/11/2016 Yes Yes Yes No 12.00 8 12/18/2016 Yes No No Yes 4.00 7 12/25/2016 No No No No 0.00 0 1/1/2017 No No Yes Yes 4.50 3 1/8/2017 Yes Yes Yes Yes 7.00 5 1/15/2017 Yes Yes Yes Yes 6.25 7 1/22/2017 Yes Yes No Yes 6.00 8 1/29/2017 Yes Yes Yes No 4.00 4 2/5/2017 Yes Yes Yes Yes 3.75 3
  • 38. Date Alice Bob Chuck Eve Length Encounters 1/1/2017 No No Yes Yes 4.50 3 1/8/2017 Yes Yes Yes Yes 7.00 5 1/15/2017 Yes Yes Yes Yes 6.25 7 1/22/2017 Yes Yes No Yes 6.00 8 1/29/2017 Yes Yes Yes No 4.00 4 2/5/2017 Yes Yes Yes Yes 3.75 3 38 Asking the Question Date Alice Bob Chuck Eve Length Encounters 1/1/2017 No No Yes Yes 4.50 ? 1/8/2017 Yes Yes Yes Yes 7.00 ? 1/15/2017 Yes Yes Yes Yes 6.25 ? 1/22/2017 Yes Yes No Yes 6.00 ? 1/29/2017 Yes Yes Yes No 4.00 ? 2/5/2017 Yes Yes Yes Yes 3.75 ?
  • 39. 39 Asking the Question Date Actual Encoutners Predicted Encounters Change 1/1/2017 3 5.0346 2.0346 1/8/2017 5 8.391 3.391 1/15/2017 7 11.7474 4.7474 1/22/2017 8 13.4256 5.4256 1/29/2017 4 6.7128 2.7128 2/5/2017 3 5.0346 2.0346
  • 42. Who is this Guy? Guy Royse Developer Evangelist Nexosis guy@nexosis.com @guyroyse 42
  • 43. 43 Credits o Images o https://www.flickr.com/photos/t17emma/5909350021 o https://www.flickr.com/photos/rueink/35095791331 o https://www.flickr.com/photos/rueink/37125691160 o https://www.flickr.com/photos/fauxlaroid/5988373053 o https://www.flickr.com/photos/kimonomania/602858091 o https://www.flickr.com/photos/wonderferret/174314534 o https://www.flickr.com/photos/herval/4100847043