From A-Life Agents to a Kingdom of N Queens IAT

1
From A-Life Agents to a Kingdom of N
QueensIAT99 Nominated to Best paper award

• Han Jing
• University of Science Technology of China
• Jiming Liu
• Hong Kong Baptist University
• Cai Qingsheng
• University of Science Technology of China

2
Outline of this talk

• 1.Introduction
• 2.The Basic A-Life Idea for Solving N-Queen
  Problem
• 3.The A-Life Agent Model for N-Queen Problem
• 4.Experimentation
• 5.Discussion and Conclusion
• 6.Future Work

3
1. Introduction

• Model from the nature Living environment
  Individual agent Interacting rules
• Domain Constraint Satisfaction Problems(CSPs)An
  Example N-Queen Problem(NQP) a newly explored
  area of research
• Goal

Environment Distributed agents Reactive rules
? Intelligence
4
1. Introduction What is a CSPs solver?
Given

• Variable set XX1,X2,,Xn
• Domain set DD1,D2,,Dn (Xi?Di )
• Constraint set RC1,C2,,Cm

An solution Assignment of values satisfying
all constraints
5
An Example N-Queen Problem(NQP)

• X n indistinguishable queens
• D an NN chessboard
• R no collisions no two queens placed on
• the same row
• the same column
• or the same diagonal
• An NP hard problem

6
2.The Basic A-Life Idea for Solving NQP
Environment Distributed agents Reactive rules
? Intelligence
Naïve Construction

• Chessboard ? Environment
• Queen ? Agent
• Constraint ? Agents moving strategy (move to a
  non-collision position)

Nothing but the traditional search!!!
7
2.The Basic A-Life Idea for Solving NQP
Survival of the Fittest High-fitness agent ?
Gain energy ? Survival (form a solution)
Low-fitness agent ? Punished, lose energy ? Die
Environment Distributed agents Reactive rules
? Emergent intelligence
The New Construction
Easy to find a position! MOVE ?SELECTION

• Environment Chessboard dynamic, recording the
  number of current collisions.
• Agent Queen energy, easy moving strategies
  (random move, move to the least collision
  position)
• ModelSurvival of the Fittest (energy
  loss,eat,die)
• Potential power to find a solution avoid losing
  energy

8
3.The A-Life Agent Model for NQP
Agent
Collision number

• Initial Energy agi.energye0
• Moveright/left
• Lose energy (?energy)
• each move (?energy1 unit)
• move to a lattice with collision number m
  (?energym units)
• Dieenergyltthreshold (suppose to be zero)
• Eatag1 meets ag2
• if ag1. energy-ag2.energy gt Merge Threshold
• then ag2 die, ag1 get ag2.energy
• Moving strategies (different probability)
• Random-move
• Least-move least collision number lattice
• Coop-move cooperating with some agents

9
E10

die
eat
8
7
3
4
least-move
9
3.The A-Life Agent Model for NQPSystem Algorithm
Initialization
Yes
No
Dispatch agents
Is a Solution ?
Yes
After some low-fitness agent died, the
systemwill be more efficient.
Output
Wanted AnotherSolution ?
10
Environment

• NN square lattice. Each lattice records
  1.what agents are on it? 2.the number of
  collisions

8
6
4
10
2 circles means 2 agents
4
8
10
11
9
12
5
6
9 agents conflict this positionDeeper blue color
means more collisions there
9
6
9
9
11
A Snapshot of the System

• Initialization
• Initial agents energy, different strategy
  parameters
• M agents/row (mgt1)
• Randomly placed

Prandom-move 0.5 Pleast-move 0.4 Pcoop-move 0.1
Lattice Deeper color means larger collision
number
gt
Prandom-move 0.05 Pleast-move 0.8 Pcoop-move
0.15
T0
12
A Snapshot of the System
Operation At each time step dispatch each
agent Reaction lose energy, die, eat
13
A Snapshot of the System
Pervious solution
T9
14
4.Experimentation - An case study

• 1500-queen (limited by the memory) CPU P233,
  RAM32M, OS Win95.

Runtime(s)
N of Queens
15
4.Experimentation
? Observation Chaos in the system
Exp1 Prandom-move0.01, Exp2 Prandom-move0.02
Prandom-move is high ? efficiency
decreases Prandom-move0 ? system falls into a
local optimum
16
? Observation1survival of the fittest
4.Experimentation
Exp1N(20,410),RowNum1,RunNum1,MaxRandom-p1,
MaxLeast-P90, MaxCoop-P20 Exp2N(20,410),RowNum
2,RunNum2,MaxRandom-p1, MaxLeast-P90,
MaxCoop-P20 Exp3N(20,410),RowNum3,RunNum3,
MaxRandom-p1, MaxLeast-P90,MaxCoop-P20
More agents ? More choice !!!
Bad agent ? punish ? die ?system efficiency
increase!
17
4.Experimentation
? ObservationPLeast-move vs. PCoop-move
PLeast-move is more important than PCoop-move
18
5.Discussion and Conclusion
Environment Distributed agents Reactive rules
? Intelligence

• Better than Network GA
• Distributed and no centralized control
• Agent selection Survival of the Fittest

A-Life1500-queen (several hundred seconds)
(Traditional search 96-queen)Hopfield/
Minimum Network 100-queen GA 200-queen
19
Lets see the demo --Readme of the demo

• For example, if you like to see how use 3
  agents/row to solve 40-queen problem
• 1.click Settings in the main menu
• 2.write 40 in the Queen Num item, write 3 in the
  AgentNum/Row item
• 3.click OK of the dialogue box
• 4.click Newgame-gtRandom in the main menu
• 5.click Resume in the main menu to see the
  running of the system, if you want to pause it,
• you can click Pause
• Or click Step in the main menu if you want to
  watch how it runs carefully
• 6.when it find a solution, a message box will pop
  up, click on it
• 7.Another message box pop up, click on it, then
  the system will randomly place the survival
• agents and continue to find another solution.
• Goto 5.

20
Run it by clicking here!
21
6.Future Work

• Solve an (Nm)-Queen Problem based on solving an
  N-Queen Problem
• Introduce reproduction and mutation
• Utilize the A-Life model to solve other CSPs

Thank You!

• CSPsX(variable),D(domain),R(constraint)
  A-Life System.
• X Agent (m agents represent one variable,
  mgt1)
• D R Environment Rules
• Solution Positions of the current survival
  agents

22
This PowerPoint file and demo www.comp.hkbu.edu.hk
/hanjing/nq.html