Monkeys in the Machine

Can a million monkeys given a million typewriters produce the collected works of Shakespeare? Considering the ethical concerns of forcing monkeys to write, paired with the sheer mess of that many mammals in one location, we might never know. Or will we?

Thanks to the wonders of modern computing, it's possible to _simulate_ monkeys emulating the Bard. Leveraging techniques like genetic algorithms, we can build a system that approximates a monkey armed with a typewriter. Digitally.

It's a fun exercise to prove a thought experiment, but also a powerful technical approach to leverage machine learning and rudimentary AI in various places. Together, we'll breed digital monkeys and show both how they can be used to answer the aforementioned Shakespearean question as well as produce generative results to more _practical_ problems.

Transcript

1. 1
   Monkeys in the Machine
   Eric Mann
   

   View Slide

2. 2
   Hello.
   Eric Mann
   Director of Engineering, Vacasa
   

   View Slide

3. 3
   Concevons qu'on ait dressé un million de singes à frapper au hasard sur les
   touches d'une machine à écrire et que […] ces singes dactylographes travaillent
   avec ardeur dix heures par jour avec un million de machines à écrire de types
   variés. […] Au bout d'un an, [leurs] volumes se trouveraient renfermer la copie
   exacte des livres de toute nature et de toutes langues conservés dans les plus
   riches bibliothèques du monde.
   — Émile Borel
   

   View Slide

4. 4
   Let us imagine that a million monkeys have been trained to strike at random on
   the keys of a typewriter and that […] these typing monkeys work hard ten hours
   a day with a million typewriters of various types. […] At the end of a year, [their]
   volumes would find themselves containing the exact copy of books of all kinds
   and all languages kept in the richest libraries in the world.
   — Émile Borel (via Google Translate)
   

   View Slide

5. View Slide

6. 6
   CREATE
   POPULATION
   EVALUATE
   FITNESS
   MUTATE REPRODUCE
   Randomly-initialized strings can eventually converge to meet a target string over
   several iterations of a random optimization algorithm over a large set of candidates.
   Evaluate how close each
   candidate is to the optimum
   Evaluate Fitness
   Initialize a collection of
   pseudrandom candidates
   Create Population
   Create a new candidate for
   the next generation
   Reproduce
   Mutate the new candidate so
   we can further optimize
   Mutate
   Genetic Algorithms
   

   View Slide

7. 7
   Image by GeeksforGeeks
   

   View Slide

8. 8
   PHP Demo
   

   View Slide

9. 9
   Genetic Algorithms
   In Review
   1
   We created a collection of entirely random strings, all the same length as our target text.
   Initialize Population
   2
   We measured the distance of each string from the target text.
   Evaluate Fitness
   3
   We used our “high fitness” members to create a new, somewhat less random child generation
   Reproduce
   4
   We mutated some children to introduce randomness and prevent local maxima.
   Mutate
   

   View Slide

10. 10
    The Traveling Salesman Problem
    Salesman must travel to all cities once before returning home.
    The distance between each city is given and is assumed to be the
    same in either direction.
    OBJECTIVE - Minimize the total distance traveled on the trip.
    

    View Slide

11. 11
    Wikimedia Commons
    

    View Slide

12. 12
    Genetic Algorithms
    Traveling Salesman
    1
    Create a “population” of randomly-ordered strings representing a route.
    Initialize Population
    2
    Measure the total distance traveled while traversing each route.
    Evaluate Fitness
    3
    We used our “high fitness” members to create a new, somewhat less random child generation
    Reproduce
    4
    Mutate by swapping the order of two cities in the route.
    Mutate
    

    View Slide

13. 13
    PHP Demo
    

    View Slide

14. 14
    Wikimedia Commons
    

    View Slide

15. 15
    Ant Colony Optimization
    Traveling Salesman
    1
    Drop an “ant” at a random city.
    Random Starting Point
    2
    The ant visits each city in turn at random, weighted by the attractiveness of each destination.
    “Attractiveness”-based Route
    3
    An ant deposits pheromone on each edge it travels, with an amount inversely proportional to
    the total distance travelled in its route.
    Calculate Pheromone Trail
    4
    The total pheromone for the next iteration is part left over from the first and part added to by
    the ants - repeat all steps until we converge.
    Evaporate and Repeat
    

    View Slide

16. 16
    PHP Demo
    

    View Slide

17. 17
    Ant Colony Optimization
    Traveling Salesman
    Vehicle Routing
    Job Ordering
    “Knapsack Problem”
    Nanoelectronics Design
    Image Processing
    Antenna Optimization
    Network Routing
    Protein Folding
    

    View Slide

18. 18
    Questions?
    

    View Slide

19. 19
    Thank You!
    eric.mann@vacasa.com | 503.925.6266
    

    View Slide