packages icon
BUGWORLD - Copyright 1993 Martyn Amos

1. INTRODUCTION

BugWorld is a simple introduction to the world of Genetic Algorithms
(GAs). A virtual physical environment is created within the computer,
which is then populated by various creatures, or "bugs", composed entirely
of machine instructions (cf. Tierra).

The environment may be visualized as a 2D, toroidal grid, the size of which
may be specified at the start of each simulation. At the start of each
generation, the bugs are randomly distributed over the grid.

The bugs compete for both energy and material resources. The
energy resource is represented by food items distributed around the
environment. Bugs may eat this food, which increases their energy 
level. Bugs lose energy at a constant rate, and must occasionally replace
it. If a bug's energy level reaches zero, it becomes exhausted and must rest
for a certain period of time, after which its energy level is restored to
its initial value.

The material resource is represented by items of litter. The more litter
items a bug (litterbug, geddit?!) collects, the more successful it is
deemed to be. Both resources are also randomly distributed at the start
of each generation. Food and litter items are generated anew (ie. each
generations is faced with a brand-new, unique environment).

2. THE POPULATION

The genome of the population members consists of a sequence of machine
instructions that encode the organism's behaviour. There are two
species of organism - LitterBugs and Bullies. LitterBugs must move around
the environment, foraging for food and collecting litter. At the end of
each generation, certain LitterBugs are selected as parent of the next
generation.

Bullies are predators in that their sole task is to attack LitterBugs and
deplete their litter collections. Bullies do not require food and do not
breed - they have a fixed genome.

At the start of each simulation, each LitterBug has a pseudo-random genome
assigned to it.

3. THE VIRTUAL COMPUTER

Each organism has associated with it a program counter (PC) and
several registers/flags:

	- memory register (integer)
	- energy level register (integer)
	- current direction register ([N|S|E|W])
	- resting register (integer)
	- litter collected register (integer)
	- organism type flag ([L|B])
	- exhausted flag (Boolean)

4. THE BUGWORLD LANGUAGE

The complete list of instructions is given below:

Opcode	Mnemonic	Interpretation

00	LOOK		"Look" in the current direction. A value
			corresponding to what can be seen is placed in
			the memory register:

				0: nothing
				1: food
				2: litter
				3: LitterBug
				4: Bully

01	MEMUP		Increment value in memory register
02	MEMDOWN		Decrement value in memory register
03	RESTART		Jump to start of program (ie. PC = 0)
04	MOVE x		Move forward with an x% probability
05	TURN x		Turn left or right with an x% probability
06	JUMP x		Jump to program line x
07	REMEMBER x	Place x in memory register
08	REST x		Rest for x cycles (consume no energy)
09	ENERGY x y	If energy < x, jump to program line y
10	LESS x y	If value in memory < x, jump to program line y
11	MORE x y	If value in memory > x, jump to program line y
12	EQUAL x y	If value in memory = x, jump to program line y

Each gene (ie. program line)  consists of three components, which are 
copied "en bloc". It is important to note that when an organism's genome is
first created it is only pseudo-random, in that the different fields of the
genes must remain within their defined limits. For example, the gene
04 103 ??? would be illegal, as it corresponds to MOVE 103%. This is
clearly illegal.

The Bully's fixed genome corresponds to the program below:

	0	LOOK		(in the current direction)
	1	EQUAL 3 5	(is LitterBug visible?)
	2	TURN 75%	(didn't jump, so try to turn)
	3	MOVE 75%	(move in another direction)
	4 	RESTART		(and then look again...)
	5	MOVE 100%	(move towards LitterBug)
	6	RESTART		(continue this process...)

5. THE SELECTION MECHANISM

Several selection schemes are implemented. The choice of mechanism can
greatly affect the rate of evolution (ie. how quickly a reasonably efficient
LitterBug evolves).

5.1 Random Selection

It is perhaps difficult to justify calling random selection of parents a
selection "strategy" at all. It simply involves choosing two members of the
population, purely at random, as parents. Each organism, regardless of its
fitness has an equal chance of selection. This mechanism is employed
mainly to demonstrate that other strategies are more effective.

5.2 Roulette Wheel Selection

The most commonly used strategy used in GA applications is probably stochastic
selection with replacement (or "roulette wheel" selection). We define the
probability that organism x is chosen as a parent as

                f(x)
	P(x) =  ----
                f(p)

where f(x) is organism's fitness
      f(p) is total fitness of entire population

This method is called "roulette wheel" because each individual is assigned
a wedge of an imaginary roulette wheel, the size of the wedge being     
proportional to the individual's fitness taken as a proportion of the total
fitness of the entire population. The greater an individual's fitness,
the larger the likelihood that the roulette ball will stop within that
individual's wedge, thus selecting it as a parent.

"Pure" roulette wheel selection uses this method to select both parents.

5.3 Impure Roulette Wheel Selection

This uses a combination of pure roulette wheel and random selection, one 
parent being selected by each method.

5.4 Tournament Selection

Each parent is selected by selecting two individuals at random and comparing 
their fitnesses. The probability that the fittest individual "wins" the
encounter and is selected as a parent is usually set to 75%. This procedure
is performed twice, once to select each parent.

6. FITNESS MEASUREMENT

So far we have assumed that a fitness measurement is available, without
discussing how it may be obtained. A reasonable measure is the amount of
litter that a LitterBug collects within a single generation, so we will
use that.

7. REPRODUCTION

Reproduction involves taking two parent LitterBugs and combining their
genomes to create an offspring's genome. When this occurs, a random section
of one parent's genome is joined to a section of the other parent's
genome, a process known as crossover.

For example, if parent A's genome is

	##########

and parent B's genome is 

 	XXXXXXXXXX

with a totally arbitrary crossover point denoted by |, we get

	##########
           | 		= ####XXXXXX
	XXXXXXXXXX

as the offspring's genome.

8. MUTATION

While copying instructions during reproduction, genes are randomly altered
at some rate. The copy mutation rate is set to an artificially high
value (1%) in order to illustrate the concept.

Usually, mutation involves "flipping" one bit of the genome, but because
of the brittle nature of the genome, flipping a bit within a gene may
result in an invalid instruction. Therefore, mutation is (very) loosely
emulated by randomly creating a whole new gene and substituting it for the
one selected for mutation.

10. BUGWORLD OPERATION AND CONFIGURATION

The program must be invoked from the command line, using the following
syntax:

	bugworld -c <configuration filename> [-t]

The -t flag is used to select terse statistics.

A lot of application parameters may have their values specified at the
start of each run. For ease of use, the starting configuration must be read
in from a text file. This file contains an arbitrary number of entries of
the form

	<parameter name> <value> <NEWLINE>

The available parameter names and their associated range of values are
described below:

Parameter name		Description		Range

POPSIZE			Size of initial		Even number 2..200
			population

GENERATIONS		No. of generations to 	1..1000
			simulate

GENOMESIZE		Size of LitterBug 	7..25
			genome (no. of genes)

BUGVISION		LitterBug's range of	1..20
			vision

BULLYVISION		As above, for Bullies	1..20

SEED			Random number seed	Any integer

XBOUNDARY		X dimension of world	50..500

YBOUNDARY		Y dimension of world	20..500

FOODAMOUNT		Amount of food at	100..5000
			start of generation

LITTERAMOUNT		As above, for litter	100..5000

SELECTION		Selection mechanism	1..4

MUTATIONRATE		Probability that a	0..100
			gene mutates when copied

MOVES			Moves per organism per 	1500..10000
			generation

FITTESTCHANCE		Probability of fittest	0..100
			organism being selected
			(tournament only)

TRACE			Number of organism to 	As POPSIZE
			trace

Default values are given below:

Parameter name		Default value

POPSIZE			UNDEFINED
GENERATIONS		UNDEFINED
GENOMESIZE		15
BUGVISION		5
BULLYVISION		5
SEED			Current process' PID
XBOUNDARY		150
YBOUNDARY		60
FOODAMOUNT		300
LITTERAMOUNT		900
SELECTION		3 (Impure roulette)
MUTATIONRATE		1
MOVES			3000
FITTESTCHANCE		75
TRACE			UNDEFINED

POPSIZE and GENERATIONS must be defined by the user at all times.
An example configuration file is given in the ir.cfg file.

11. STATISTICS AND TRACING

The user of the package may select the type of statistics displayed
during a simulation. If terse statistics are selected, the program
simply prints out the average fitness of the population at the end of
each generation.

If terse statistics are not selected (the default), the program displays
a header, and various statistics for each generation (generation number,
average fitness, number and fitness of fittest LitterBug and the genome
of the fittest LitterBug).

All statistics are printed to standard output (usually the terminal
screen). This may be redirected to a text file or piped to a printer.

The facility to trace the execution of an individual organism's genetic
program is available. This is mainly used for debugging purposes, but
is useful for watching an organism's interactions with its environment.
The number of the organism to be traced must be specified by the
TRACE parameter. The organism is traced for each generation and the
trace information is placed in a file named 'trace' in the current
directory.

12. CONCLUSIONS

The behaviour of the simulated population generally conforms to basic
principles of GAs, such as premature convergence. It was hoped that
"intelligent" LitterBugs that avoided predators and efficiently foraged
for food would evolve, but this didn't happen during the limited number of
runs performed, due to time constraints.

The package doesn't try to stick to formal GA principles (for example,
mutation is performed in a very crude manner), but it does illustrate
several basic concepts, such as crossover, selection and convergence.

Comments regarding this software are welcome. Please direct correspondence
to

	Martyn Amos
	Parallel Computation Group
	Department of Computer Science
	University of Warwick
	COVENTRY
	United Kingdom
	CV4 7AL

	martyn@dcs.warwick.ac.uk