a case study of on-the-fly calibration for artificial ...ls11- fileoutline 1 introduction 2 arti...

Introduction Artificial Immune Systems Proposal Experiments and Results

A Case Study of On-The-Fly Calibration for Artificial ImmuneNetwork Algorithms

Elizabeth MonteroUniversidad Tecnica Federico Santa Marıa

Joint Workshop on Automated Selection and Tuning of Algorithms,

Part B) Discrete Search Spaces - Focus on Parameter Selection

1/19


Introduction

Outline

1 Introduction

2 Artificial Immune Systems

3 Proposal

4 Experiments and Results

2/19


Introduction

Motivation

Motivation

To extend the adaptive approach presented for clonal selection immunealgorithms to a sophisticated immune network algorithm that solveshard constraint satisfaction problems

QuestionsI If tuning process selects different values for some parameters when

tuned for different instances

I then there is possible to detect the need of controlling the moresensible parameters of the algorithm?

I It is possible to reduce the number of parameters of an algorithm bycontrolling some of them?

I Incorporates parameter control to an algorithm can improve itsefficiency?

3/19


Artificial Immune Systems

Outline

1 Introduction


3 Proposal


4/19




Artificial Immune SystemsI A new bio-inspired approach to Artificial Intelligence

I Based on the capabilities of the adaptive immune system of vertebrateorganisms

I UniquenessI Extern recognitionI Anomaly detectionI Distributed detectionI Imperfect detectionI Reinforcement learning and memory

I Used in optimization and machine learning

5/19



Artificial Immune Processes

Artificial Immune ProcessesI Positive/Negative Selection

I T-cells determination

I Clonal Selection process

I Infection response

I Artificial immune networks

I Balance methods

I Danger Theory

I Danger detection in organisms

6/19



Artificial Immune Networks

7/19



CD-Nais parameters

B Number of cells in repertoiren Number of cells selected for cloningC Number of clones generated from cellsr Mutation rate (Mutation rates)u Network interactions thresholdd Number of new random cells incorporated each iteration

8/19


Proposal

Outline

1 Introduction


3 Proposal


9/19


Proposal

Our proposal

IdeaI Award good actions and punish bad actions

I balance diversification and intensification of search process

What to control?I Diversity: Number of selected cells (n)

I Decrease the number of selected cells implies to increase the amountof random cells incorporated to population

I Intensification: Number of clones (C)

I Increase the amount of clones generated by each selected cells impliesto increase the intensification level

10/19


Proposal

Adaptive CD-Nais algorithm

11/19


Proposal

Adaptive CD-Nais algorithm

12/19


Proposal

In summary

SummaryI Improvement due to mutation procedure implies an increase of

Intensification process.

I Deterioration in the whole population quality, implies a reduction ofthe amount of random antibodies incorporated to the population,reducing the diversity of the population.

13/19


Experiments and Results

Outline

1 Introduction


3 Proposal


14/19



Constraint Satisfaction Problems

I Classical benchmark problems

I Problem:

I Set of VariablesI Set of DomainsI Set of ConstraintsI Objective: To find an assignment of values to variables that satisfies all

constraints

I 3-coloring problem instances

I 30, 45, 60, 75, 90, 105, 120, 180 vars

15/19



Tests on 3-Coloring problems

I Controlled version shown a better performance than static versionspecially in largest instances.

I Further analysis make us suspect about the irrelevance of parameter (u) inadaptive CD-Nais

Final Survivors:params/configuracion8.param 0.0625 0.0566733params/configuracion0.param 0.135417 0.0989402params/configuracion10.param 0.135417 0.0989402params/configuracion1.param 0.125 0.0882764params/configuracion2.param 0.125 0.0833682params/configuracion3.param 0.145833 0.100459params/configuracion4.param 0.135417 0.0900236params/configuracion5.param 0.114583 0.0911859params/configuracion6.param 0.114583 0.070321params/configuracion7.param 0.104167 0.0679408params/configuracion9.param 0.104167 0.0679408Total Experiments: 1056

Table: F-Race results

16/19


AnswersI If tuning process selects different values for some parameters when

tuned for different instances then there is possible to detect the needof controlling the more sensible parameters of the algorithm?

I Yes, in most cases.

I It is possible to reduce the number of parameters of an algorithm bycontrolling some of them?

I Controlling parameter values can produce direct and indirect effectover other parameters of the algorithm allowing the reduction of thetotal number of parameters.

I Incorporates parameter control to an algorithm can improve itsefficiency?

I Yes, it can reduces the effort spent in tuning processes.

17/19



Questions?

Any question?

18/19



References I

Hutter, F., Hoos, H.H., Stutzle, T.: Automatic algorithm configuration based on localsearch. In: Proceedings of the Conference on Artificial Intelligence. pp. 1152–1157(2007)

Riff, M.C., Montero, E.: A dynamic adaptive calibration of the clonalg immunealgorithm. In: Proceedings of the International Conference on Adaptive and IntelligentSystems. pp. 187–193. IEEE (Sep 2009)

Riff, M.C., Zuniga, M., Montero, E.: A Graph-based Immune-inspired ConstraintSatisfaction search. Neural Computing & Applications pp. 1–10 (2010)

19/19

a case study of on-the-fly calibration for artificial ...ls11- fileoutline 1 introduction 2 arti...

Documents