CATEGORIZING COREWAR WARRIORS

1
CATEGORIZING COREWAR WARRIORS

• Nenad Tomaev, Doni Pracner

2
Data mining

• A process of extracting non-trivial, previously
  unknown and potentially useful pieces of
  information from a set of data
• Main types
• Classification
• Association
• Clustering
• Numeric prediction

3
COREWAR

• Assembly language redcode
• Programs compete for resources
• Virtual world where the battle takes place Core
• MARS Memory Array Redcode Simulator

4
(No Transcript)
5
COREWAR

• Great variety of strategies
• Replicators
• Scanners
• Coreclears
• Stones
• Vampires
• Hybrid strategies

6
EVOLUTIONIN COREWAR

• Warriors generated via genetic algorithms
• Fitness a) benchmark score
• Fitness b) round robin tournament score
• Coevolution, Island model

7
PROJECT GOAL

• Examining diversity among the warrior set
  generated by CCAI evolver via the use of
  clustering algorithms

8
CCAI EVOLVER(Barkley Vowk, Canada)

• Island model
• 20 random species
• Pool for the next generation is the previous
  generation

9
THE DATASET

• 26795 warrior files
• Chronoligically divided into 4 groups
• 10554
• 6889
• 4973
• 4689

10
DATAREPRESENTATION (1)

• Only command names counted
• Modified bag of instructions
• ADD and SUB joined
• Presence or absence of imp components
• Characteristic consecutive instruction pairs

11
The Application Created

• C\data mine\gtjava Worker
• CoreWar analyzer
• Usage
• Worker wu indir outdir outfile- complete
  process
• Worker w indir outdir - just analyze and create
  the files
• Worker u indir outfile - unite the files in
  the dir into a csv
• Worker d indir outdir adir - check for
  duplicates in indir.
• adir should be made by w
• Worker s indir outdir adir - (very) speeded up
  version of d
• C\data mine\gt

12
DATAREPRESENTATION (2)

• 30 warriors benchmark selection
• Score table for each set was created

13
FILTERING DATA

• Removing duplicates in the representation space
• Motivation
• Speeding up the process
• Ensuring better clustering algorithm performance
• Duplicate any two warriors with identical
  instructions and modifiers, ignoring address
  fields

14
FILTERING DATA (2)

• Dataset4 4389 warriors
• 43894388/2 9.629.466 about 100 lines, each 4
  parts to compare
• about 3.851.786.400 logical comparisons.
• Method duplicates only 25.309 file comparisons
  188 minutes
• Method duplicatesSeparation dataset 3 4973
  warriors 29 minutes

15
FILTERING DATA (3)results

• 2150 instances were removed (8) or per part
• 12 - 1345
• 8 - 559
• 1 - 56
• 4 - 193

16
Clustering

• EM clustering method was used (expectation
  maximization algorithm)
• Weka (Waikato Environment for Knowledge Analysis)

17
RESULTS
18
RESULTSAttribute Evaluation
19
RESULTSAttribute Evaluation
20
RESULTSAttribute Evaluation
21
RESULTSAttribute Evaluation
22
RESULTSAttribute Evaluation
23
GROUP 1 (JMN)
24
GROUP 1 (SPL/MOV)
25
GROUP 2 (DAT)
26
GROUP 3 (SPL)
27
GROUP 4 (DJN)
28
SOME CONCLUSIONS

• Over time, the apparent loss of diversity was
  recorded, due to the survival of the most
  efficient mutation-resistent forms
• Replicators and coreclears dominated the
  population
• Presence of imps was noted in a large part of the
  population

29
CONCLUSIONSAttribute evaluation

• Attribute evaluation did notice big differences
  between the original pool (set 1) and the first
  generation (dataset 2)
• DJN and MOVDJN were the most significant
  attributes in the clustering of the whole data
  set.
• SPLMOV and MOVJMP were also important in
  clustering of some of the subgroups.

30
FUTURE RESEARCHIDEAS

• Use of different clustering methods, results
  comparison
• Training the classifiers for identification of
  human-coded warriors
• Based on syntax analysis
• Based on benchmark score

31
THE END

• Thank You for Your attention