Using a Genetic Algorithm to Create Prey Tactics

1
Using a Genetic Algorithm to Create Prey Tactics

• Presented by Tony Morelli on 11/29/04

2
Abstract

• Evolve a predator/prey system
• Inputs are the distances/angles between each
  predator/prey and other obstacles
• Outputs are the bearing and speed to avoid
  confrontations while maintaining a task.
• Focus on a prey whose mission is to follow the
  shoreline.

3
Introduction

• Use a Genetic Algorithm to evolve a prey tactic
  when being attacked.
• This example uses animals placed inside the SWARM
  architecture.
• Objective for the prey is to identify a predator
  and avoid contact with it.
• Prey behaviors/actions will be evolved against a
  hand coded predator.

4
Introduction

• Evolving predator/prey is useful because it can
  demonstrate what happens in nature as well as
  create new military tactics.
• GA is useful because we have a set of behaviors
  and a set of triggers for those behaviors. Hand
  coding this is difficult. A GA should find the
  best set of behaviors and triggers.

5
Background

• Bauson Ziemke
• Evolved View Angle, View Range
• Speed was used as a constraint
• Prey prefers a camera with wide angle and short
  range, while a predator prefers small angle and
  long range
• Predators dominated prey

6
Results Summary

• Evolved prey outperformed hand coded prey when
  placed against a hand coded predator.
• Evolved predator outperformed hand coded predator
  when placed against hand coded prey
• Evolved predator outperformed evolved prey

7
Introduction

• Methodology
• Results and Analysis
• Conclusions/Future Work

8
Methodology

• Prey needs to know it is being attacked and then
  react to the situation
• Preys primary goal is to follow the shoreline
  clock-wise
• Must avoid predators and land

Follow Shoreline
9
Methodology

• Genetic Algorithm was used to evolve prey tactics
• GA by Ryan Leigh
• 1 Point Crossover
• Elitist Selection
• Crossover 0.7
• Mutation 0.1
• Population 20
• Generations 20

10
Methodology

• Parameters
• Distance from predator
• Far, Near, Close, TooClose
• 50-944 pixels
• Speed
• Slow, Normal, Fast, Superfast
• 0.025-0.3
• Turning Rate
• p/16 p / 2 radians
• Vision Range
• p/16 p / 2 radians

11
Methodology

• 51 Bit String
• Bits 0-7 Far
• Bits 8-15 Near
• Bits 16-23 Close
• Bits 24-30 Too Close
• Bits 31-33 Turning Rate
• Bits 34-36 Vision Range
• Bits 37-43 Fast Speed
• Bits 44-50 Normal Speed

12
Methodology

• Parameter values were evolved
• When each parameter was used was not evolved
• If enemy is too close change speed to Super Fast
• The values for super fast and too close were
  evolved, not the logic surrounding them

13
Methodology

• Once an attack is identified, the prey will try
  to avoid contact.
• When anything gets within certain ranges, or a
  crash is projected within a certain range, the
  prey will react to it

14
Fitness Evaluation

• Success is measured by time
• Until the prey thinks he is being attacked,
  fitness increments by 1 every update
• If the prey is wondering around and never
  encounters a predator, his evasive skills are not
  tested, so this allows to keep that prey alive in
  the gene pool
• Once an attack is detected fitness is incremented
  by 5 every update
• We really want to measure the preys evasive
  ability. This weight allows for that.

15
Methodology

• The simulation was run for 5 minutes
• This was at an accelerated rate
• 5 minutes would take a few seconds
• If at any point the predator/prey collide, or
  either one hits land, the simulation ends
• Fitness was calculated and the GA performed its
  job.

16
Methodology

• First the default predator and the default prey
  went head to head to calculate a fitness.
• Next the prey was evolved against the default
  predator. The top prey then went head to head
  against the default predator
• The predator was evolved against the default
  prey. The top predator then went head to head
  against the default prey
• Finally the evolved predator and the evolved prey
  were matched up and the fitness of the prey was
  evaluated.

17
Results

• Default Predator vs Default Prey

Seed Fitness
0.1337 130189
0.8712 74867
0.7107 89023
0.835 67161

• Average 90310

18
Results
19
Results

• Evolved Prey vs Default Predator

Seed Fitness
0.1337 173523
0.8712 303250
0.1707 116531
0.835 205971

• Average 199819
• 221 Increase

20
Results

• Evolved Predator vs Default Prey

Seed Fitness
0.1337 22693
0.8712 50037
0.1707 41991
0.835 59181

• Average 43476
• 48 Decrease

21
Results

• Evolved Predator vs Evolved Prey

Seed Fitness
0.1337 26873
0.8712 34326
0.1707 19303
0.835 30181

• Average 27671
• 70 Decrease

22
Results

• Evolved Predator vs Evolved - Evolved Prey

Seed Fitness
0.1337 172865
0.8712 152757
0.1707 200454
0.835 249813

• Average 193972
• 214 Increase

23
Analysis

• As expected, evolved prey was highly successful
  when compared to the default predator
• Evolved predator was much better than evolved
  prey
• Evolved prey developed specialized parameters
  that were successful against 1 type of predator
• Once paired against a different predator, the
  learned tactics no longer applied
• No general knowledge

24
Conclusions

• The GA did work against a known predator
• My evolved prey did not develop any general
  knowledge

25
Future Work

• Need to add in logic for random turning when
  there is no limit on turning
• Need to add in logic for handling multiple
  predators.
• Should plan a route instead of just reacting and
  running away.