Artificial Agents
21
2. For each machine, observe an input symbol and output produced.
3. Find a method to measure outputs, using a payoff function, known as
utility function.
4. Determine fitness of each machine depending on result of the utility
function.
5. Machines with a higher payoff are retained to be parents for the next
generation of machines. Sometimes half the population is retained until
next iterative step.
6. Offspring or new machines can be produced as combinations of two
parent machines or by mutation (varying an input symbol or a next
state).
Fogel [61] modified complete state machines using evolutionary programs,
such as using state machines to play the prisoner dilemma games. These machines were represented as string structures, where genetic operations like
crossover and mutation can be applied to modify their structures.
Rechenberg [155] and Schwefel [172] viewed genes as behavioral traits of
individuals. Their evolutionary strategies represented the gene as a vector over
n dimensions, where mutation and crossover can be performed on n dimensions and on the vector. Holland [88] [89] proposed using genetic algorithm
as a search method for adaptive systems. All of these methods use machines
to track a particular fitness landscape in a domain to find how far the machine is from ideal. Therefore, this highlights that genetic algorithms in agent
architectures would require fitness landscapes to work with.
Put forth as a ‘thought experiment’, von Neumann [198] presented a hypothetical model of a machine that used raw materials from the environment
to produce a second machine by replicating itself. Self-replicating automaton
presented the grounds for building cellular automata experiments triggering
research in AI, where geography and interactions influenced the production
of new machines [105]. Although such a self-replicating robot in the physical
world may still be a budding area of research, the concept was introduced
and tested in a virtual world of simulation, extended using cellular automata,
later giving birth to agent-based modeling methods.
2.2 Engineering Self-Organizing Systems
A model is an approximate representation of a system, showing only basic functionalities or just parts being investigated. Various systems in nature
are observed and adopted to create self-organizing systems. Insect colonies,
cells and human societies are all examples of these using stigmergy or similar
21
2. For each machine, observe an input symbol and output produced.
3. Find a method to measure outputs, using a payoff function, known as
utility function.
4. Determine fitness of each machine depending on result of the utility
function.
5. Machines with a higher payoff are retained to be parents for the next
generation of machines. Sometimes half the population is retained until
next iterative step.
6. Offspring or new machines can be produced as combinations of two
parent machines or by mutation (varying an input symbol or a next
state).
Fogel [61] modified complete state machines using evolutionary programs,
such as using state machines to play the prisoner dilemma games. These machines were represented as string structures, where genetic operations like
crossover and mutation can be applied to modify their structures.
Rechenberg [155] and Schwefel [172] viewed genes as behavioral traits of
individuals. Their evolutionary strategies represented the gene as a vector over
n dimensions, where mutation and crossover can be performed on n dimensions and on the vector. Holland [88] [89] proposed using genetic algorithm
as a search method for adaptive systems. All of these methods use machines
to track a particular fitness landscape in a domain to find how far the machine is from ideal. Therefore, this highlights that genetic algorithms in agent
architectures would require fitness landscapes to work with.
Put forth as a ‘thought experiment’, von Neumann [198] presented a hypothetical model of a machine that used raw materials from the environment
to produce a second machine by replicating itself. Self-replicating automaton
presented the grounds for building cellular automata experiments triggering
research in AI, where geography and interactions influenced the production
of new machines [105]. Although such a self-replicating robot in the physical
world may still be a budding area of research, the concept was introduced
and tested in a virtual world of simulation, extended using cellular automata,
later giving birth to agent-based modeling methods.
2.2 Engineering Self-Organizing Systems
A model is an approximate representation of a system, showing only basic functionalities or just parts being investigated. Various systems in nature
are observed and adopted to create self-organizing systems. Insect colonies,
cells and human societies are all examples of these using stigmergy or similar