Perfect Blackjack

1
Perfect Blackjack

• Using Genetic Algorithms

Henry Balanon balanonh_at_msu.edu
John Kyritses kyritses_at_msu.edu
2
Project Purpose

• When to hit/stay in blackjack
• Write the most spectacular GA mankind has ever
  seen
• Determine the best betting method

3
Things not covered

• Determine the best betting method

4
How to win?

• We know
• The dealers face-up card
• The cards we hold
• What cards have been played
• What actions we can take
• 4. Hit or stay

5
Representation
1. The Dealers Face-up Card
2. The Players Cards
6
Representation
7
How to win?

• We know
• The dealers face-up card
• The cards we hold
• What cards have been played
• What actions we can take
• 4. Hit or stay

8
3. What Cards Have Been Played

• Running average of the discard pile
• Example
• Player has K and 10
• Dealer has Q and J
• 4 Cards played
• The average card in the discard pile is 10

9
What does this mean?

• If (average discard high)
• Low card probable
• If (average discard low)
• High card probable

10
How to win?

• We know
• The dealers face-up card
• The cards we hold
• What cards have been played
• What actions we can take
• 4. Hit or stay

11
Representation
12
Representation
13
Evaluation Function
Fitness Wins / t_Games
14
Evaluation Function

• Played Blackjack
• Players hands played by the given rules
  (chromosome)
• Dealer stayed on 17 and up
• A blackjack for the player nothing
• Played thousands of games per evaluation

15
Genetic Algorithm Application

• Which GA application to use?
• Genetic Algorithm bLAckjack Player ? GALAP
• Ability to customize
• Number of generations
• Population size
• Crossover, Mutation Rates
• Crossover parent selection method
• Percentage of population to evolve
• Stagnation detection sensitivity
• Evaluation function
• Time vs. Accuracy

16
GALAP

• Operations
• Crossover
• Mutation
• Features
• Stagnation Detection good
• averageArchive currentAvg
• Reaction to Stagnation.

17
GALAP continued..

• Order of operations
• Stagnation detection good
• GALAP reaction to stagnation not good

18
Experiment Concerns

• Testing Environment
• Ensure reasonable accuracy while completing tests
  within time constraints
• Time per evaluation

19
Initial Tests

• Crossover 0.7
• Mutation 0.05
• Pop Size 100
• Replace Lowest 0.5
• Harsh evaluation function
• Both crossover parents chosen from top 70

20
Results

• Crossover 0.7
• Mutation 0.005
• Population 200
• Replace Lowest 0.4
• Allowing lucky blackjacks as win

21
Introduce Stagnation Control

• Crossover 0.7
• Mutation Rate 0.1
• Population Size 50
• Replace Lowest 0.5
• Crossover parents selected from top 70

22
Attacking Stagnation Repeatedly

• Crossover Rate 0.7
• Mutation Rate 0.05
• Population Size 50
• Replace Lowest 0.5
• Keep track of past 21 fitness

23
Increased Population Size

• Crossover 0.7
• Mutation 0.015
• Population Size 1000
• Replace Lowest 0.4
• Running time17 hours

24
Keeping up w/ the Jones
25
Conclusions

• GALAP vs. Straight Player
• Why?
• Greater population size
• Current GALAP reaction to stagnation is useless
  at best, harmful at worst
• Representation has potential that may still be
  realized

26
Looking Ahead

• Continue running experiments with increased
  population size for more generations
• Betting strategies
• Expand and improve GALAP
• Multiple populations
• Greater ability to customize parent selection
  methods
• Modify mutation and crossover rates when
  stagnating

27
Questions?

• Answers.