A GENOME BASED VISION OF MULTI-AGENT SYSTEMS

Monica Vitali

Universit

a degli studi di Palermo, viale delle Scienze, Palermo, Italy

Massimo Cossentino, Riccardo Rizzo

CNR-ICAR Palermo, viale delle Scienze, Palermo, Italy

Salvatore Gaglio

Universit

a degli studi di Palermo and CNR-ICAR Palermo, viale delle Scienze, Palermo, Italy

Keywords:

Agent models and architectures, Multi-agent systems, adaptation.

Abstract:

A set of software agents can be programmed to provide a large but ﬁnite set of services, often deﬁned during

design phase. After an evolution of the external environment, the pre-deﬁned services could be unable to

satisfy the requested quality. In this work an agent framework is proposed capable to adapt the agents in order

to improve the quality of services provided by an agent society in correspondence with a modiﬁcation of the

external environment. These agents are based on a biologically inspired structure (genome), that deﬁnes all

their behaviors and knowledges.

1 INTRODUCTION

Agent Oriented Systems should be able to au-

tonomously adapt and deliver new services in re-

sponse to unforeseeable problems, like many living

things do. In living things this characteristic is mainly

based on genome and natural selection. We propose

to replicate this mechanism in Agent Oriented Soft-

ware and Agent Systems.

In the proposed system agents capabilities are de-

scribed in their genome and their improvement is pos-

sible by means of a Darwinian evolution. When a so-

lution to a problem is not achievable (the correspond-

ing service is not available or it does not provide the

required quality of service), several agents can re-

produce themselves thus creating a new generation

of agents that have new capabilities and can satisfy

the requirements. The consequence of this behavior

is the creation of an adaptive system (Gleizes et al.,

1999) realized through an extensive adoption of Ge-

netic Programming techniques (Koza, 1992) and au-

tomatic code generation, compilation and execution.

The focus of our attention is not on the agent it-

self, but on its genetic makeup: the genome. This ge-

netic makeup can be decomposed in different layers.

In the ﬁrst layer of the genome there are two kinds

of chromosomes: a Knowledge Chromosome which

describes knowledge about the environment and a set

of Ability Chromosomes which describe agent’s abil-

ities to interact with the world. At a deeper layer we

have genes. Each chromosome is made of genes. In

the Knowledge Chromosome, each gene describes an

element of the knowledge (predicates, concepts and

actions), while, genes within Ability Chromosomes

describe plans components.

During the designing process it is essential to de-

ﬁne and manipulate the genome. In the deﬁnition of

the genome an initial set of genes is given. This set

does not have to contain the solution but only the ele-

ments which allow the creation of the ﬁrst generation.

From this initial set, through manipulation, genes

are combined and activated originating new genomes.

These genomes have to be evaluated through an ob-

jective function for measuring their level of adapta-

tion to the required skill.

The agent adaptation procedure is represented in

Fig.1. This process is started by an agent called

CrosserAgent after a request for an unavailable ser-

vice is received. The CrosserAgent creates new gen-

erations which will contain some agents inherited

406

Vitali M., Cossentino M., Rizzo R. and Gaglio S..

A GENOME BASED VISION OF MULTI-AGENT SYSTEMS.

DOI: 10.5220/0003165204060409

In Proceedings of the 3rd International Conference on Agents and Artiﬁcial Intelligence (ICAART-2011), pages 406-409

ISBN: 978-989-8425-41-6

 2011 SCITEPRESS (Science and Technology Publications, Lda.)

Figure 1: A representation of the adaptation procedure.

GenomeAgents are represented in white. The agent selected

at the end of the procedure is marked with an “X”, while the

CrosserAgent is marked with a “C”.

from the previous generation and some new agents

obtained crossing the agents’ genomes. The num-

ber of individuals per generation is a parameter of

the adaptation process. During the adaptation process

two agents merge their chromosomes according to the

rules of genetic programming that will be discussed

later. The parents transmit their genes to the child

which evolves and gains the capability of reaching

different results. The adaptation process ends when

an agent in the current generation provides a satisfy-

ing service or if an a priori deﬁned number of gener-

ations has been reached. The ﬁttest agent is selected

and becomes a member of the society. The CrosserA-

gent notiﬁes its name to the requesting agent in order

to fulﬁll the service request. This process will be dis-

cussed more in details in section 3.

2 THE GENOME STRUCTURE

This section introduces the Genome structure shown

in form of an UML class diagram in Fig.2. This ﬁg-

ure highlights the two main parts in which the genome

can be decomposed at a logical level: knowledge and

abilities. The division is pointed out through the in-

clusion of chromosomes in two different packages.

Starting from the higher level, the genome (at the

top of Fig.2) contains all the information needed to

describe the agent; from this information a new agent

can be created.

The genome enables a set of Services which

makes explicit the functions offered by the agent to

the external environment. The genome is composed

of a KnowledgeChromosome and a set of AbilityChro-

mosomes.

The KnowledgeChromosome aggregates genes

which refer to ontological concepts (OntologyGene)

and that are specialized in three categories: (i) Con-

ceptGene: describes an instance of a concept of the

ontology; (ii) ActionGene: describes an instance of an

action of the ontology; (iii) PredicateGene: describes

an instance of a predicate of the ontology.

The AbilityChromosome is composed of a set of

node genes which describe the plan structure (Node-

Gene) and by the contents of these nodes which de-

scribe the action associated to them. Node contents

can be of three different kinds: predicate or action

genes (indicated as PredicateGene and ActionGene in

Fig.2) or other Ability Chromosomes. Plugging in an

Ability Chromosome with a node allows us to asso-

ciate a behavior, described by another plan, to a node,

thus creating a sort of recursive structure. There are

four kinds of nodes: StartNodeGene, EndNodeGene,

ActionNodeGene and IfNodeGene. The node classi-

ﬁcation reported here is inspired by (van Der Aalst

et al., 2003).

3 THE AGENT ADAPTATION

PROCESS

The agent adaptation process, led by the CrosserA-

gent, is composed of the following iterative steps:

• deﬁnition of the parents’ sub-society: this sub-

society includes all the agents that will be used

in the adaptation process and that contribute with

their ability genes and knowledge genes to the

deﬁnition of the resulting agent;

• creation of the new generation by using adaptation

techniques;

• evaluation of the results provided by the new

agents;

• stop of the process if one (or more) agent(s) suc-

cessfully provide the required service.

During the agent adaptation process we create

new generations by using mutation, elitism and cross-

ing techniques. While the ﬁrst two techniques

are reused from litterature (Banzhaf, 1998)(Mitchell,

1998), crossing is described below.

The agent adaptation process can be divided into

two steps: knowledge crossing and ability crossing.

Knowledge crossing allows to modify the set of

knowledge genes about the environment. Knowl-

edge Chromosomes crossing is inspired by (Noy and

Musen, 1999) and it is performed over each agent’s

knowledge gene by using four techniques:

• Fusion: the two parents’ knowledge genes are

melted in a single gene;

• Selection: one of the parents’ knowledges is cho-

sen while the other one is discarded;

• Union: both of the parents’ knowledges are

copied in the new individual;

A GENOME BASED VISION OF MULTI-AGENT SYSTEMS

407

Figure 2: The genome structure. Genome is composed of two kinds of chromosome: Knowledge Chromosome and Ability

Chromosome; both are composed of genes.

Figure 3: On the left the two plans of the parent agents and

on the right the resulting plan. The parts of the plan selected

for the crossover procedure are ﬁlled.

• Copy: if a particular portion of knowledge is

present only in one of the parent, it is copied to

the generated agent.

The result is a new Knowledge Chromosome. Once

the knowledge crossing is completed, the ability

crossing can be executed.

Tha ability crossing is performed on the abili-

ties of an agent. Abilities are represented through

plans, composed of nodes, and are labeled with a goal,

which indicates the ability purpose. Agents in the

platform are provided with a higher-level plan which

handles the agent’s life-cycle and allows each agent

to interact with the external environment. This plan

is always crossed by a fusion operation. All the other

plans can be crossed also by using the selection, union

or copy techniques already described in the previous

paragraph. Since selection, union and copy simply

transfers a plan from a parent to its child, the sole op-

eration worth of a discussion is plan fusion and there-

fore it will be discussed in what follows. Fusion can

be performed only if two plans are similar (if they

have the same goal). Plan crossing is shown in Fig.3.

The two parents’ plans play a different role in this part

of the adaptation process. The receiver’s agent plan is

used as the basis for implanting the contribution from

the donor agent. An example of donor and receiver

plan fragments are shown respectively in the right and

the middle part of Fig.3. The two fragments are linked

by replacing a randomly selected node (node a’ in

Fig.3) from the receiver’s plan with the one selected

from the donor’s one (node a and its successors), as

shown in the right part of Fig.3.

4 EXPERIMENTAL RESULTS

In the case study, the initial set of agents is composed

of individuals drawing simple geometrical shapes.

The aim of each agent is to reproduce a given picture,

in the reported example a trapezium (Fig.4a).

The agent divides the picture in small chunks by

using a grid and tries to ﬁll each position of the grid

according to the guidance provided by the target pic-

ture. The work is carried on through several iterations

and each iteration is populated by a different gener-

ation of agents. The initial generation is composed

of two simple agents. Each of them is able to ﬁll

in a grid cell with the shape of a triangle: the ﬁrst

agent (Fig.4b) draws a dark gray triangle oriented to-

wards the up-left corner of the cell, the second agent

(Fig.4c) draws a light gray triangle oriented towards

the bottom-right corner of the cell.

Clearly none of two agents supply a fulﬁlling re-

sult; besides, neither the simple cooperation of the

two agents could solve the problem. So that an adap-

tation process is started: the genome codes the color,

dimension and shape of the grid cell together with the

plan used to generate the ﬁgure. The evaluation pro-

ICAART 2011 - 3rd International Conference on Agents and Artificial Intelligence

408

Figure 4: In (a) the target picture. In (b) and (c) the results

achieved by two agents of the initial society.

Figure 5: Two agents which provide the required service in

different ways.

cess compares color and shape of the obtained ﬁgure

to the target one.

In our experiments, after about nine generations,

several individuals which perfectly reproduce the de-

sired picture have been created (obviously because of

the random characteristic of the new generations pro-

duction different runs of the experiment may produce

different results). Fig.5 reports two examples of such

individuals; as it is possible to see, the two agents use

a different grid to decompose the target picture.

The test case has been evaluated with different

pictures and colors. It has been observed, as it was

expected, that the number of generations needed to

reach a perfectly ﬁtting outcome grows up with the

complexity of the target picture.

The adopted adaptation process proved to be suc-

cessful but it is to be noted that the development

framework is undoubtedly complex in its use and the

setup of a new experiment requires a lot of program-

ming.

5 CONCLUSIONS AND FUTURE

WORKS

In this paper we proposed a service adaptation mech-

anism as an integral part of an agent-oriented adaptive

and self-organizing society. As a ﬁrst step towards our

goal we tested an adaptive system inspired by the Dar-

winian evolution theory where all the agents’ features

are codiﬁed in a genome-like structure. In order to im-

prove the quality of a given service, several agents can

reproduce and generate new individuals which better

ﬁt the target. These individuals are provided with new

capabilities derived by their parents. The approach

has been tested through simple case studies. The ap-

plication reported in this paper proves that it is possi-

ble to obtain a perfectly working agent from original

agents which provides a service with a low quality

service. Using the proposed Genome Framework the

problem moves from the implementation of a solution

to the deﬁnition of the problem domain.For sure dif-

ferent techniques may be explored (and they will be

in the future) but the goal of the current study is evalu-

ating the adoption of the proposed genome-based de-

scription of agent capabilities and knowledge. The

obtained results encourage the development of further

release of the proposed framework. The use of a for-

malization language to describe the genome structure

might be the following step in order to lay the ground-

work for an agent-oriented language.

REFERENCES

Banzhaf, W. (1998). Genetic Programming: An Introduc-

tion on the Automatic Evolution of computer programs

and its Applications. Morgan Kaufmann.

Gleizes, M.-P., Camps, V., and Glize, P. (1999). A theory

of emergent computation based on cooperative self-

organization for adaptive artiﬁcial systems. In In proc.

of Fourth European Congress of Systems Science.

Koza, R. (1992). Genetic Programming: On the Program-

ming of Computers by means of Natural Selection.

MIT Press.

Mitchell, M. (1998). An introduction to genetic algorithms.

The MIT press.

Noy, N. and Musen, M. (1999). An algorithm for merging

and aligning ontologies: Automation and tool support.

In Proceedings of the Workshop on Ontology Manage-

ment at the Sixteenth National Conference on Artiﬁ-

cial Intelligence (AAAI-99), pages 1999–0799.

van Der Aalst, W., Ter Hofstede, A., Kiepuszewski, B., and

Barros, A. (2003). Workﬂow patterns. Distributed and

parallel databases, 14(1):5–51.

A GENOME BASED VISION OF MULTI-AGENT SYSTEMS

409