t uning t abu s earch s trategies via v isual d iagnosis
DESCRIPTION
T uning T abu S earch S trategies via V isual D iagnosis. >MIC 2005TRANSCRIPT
TuningTabu Search Strategies
viaVisual Diagnosis
>MIC2005<<Vienna<6th Metaheuristics International Conference
August 22-26, 2005. Vienna, Austria
By: Lau Hoong Chuin, Wan Wee Chong, and Steven Halim (Presenter)
Outline
Introduction (The Problem)– Metaheuristics Tuning Problem (illustrated using Tabu Search)
Visual Diagnosis Tuning (The Methodology)– Human + Computer– {Cause – Action – Outcome} tuple
V-MDF (Visualizer for MDF) (The Tool)– Distance Radar– V-MDF Architecture– Experimental Results
Questions & Answers
Introduction: Tuning Problem
Characteristics of a practical metaheuristic:– Delivers high quality solutions for any future instances.– Run in reasonable running time.– Can be developed within tight development time.
The need for a proper tuning.
Taxonomy of Tuning Problem:– Static versus Dynamic– Three levels of complexity
Tuning is complex… (illustrated using Tabu Search)
StaticLevel-3
TuningSearch Strategies
Level-2
ChoosingBest Configuration
Level-1
CalibratingParameter Values
DynamicTuning Problems
Tuning Tabu Search
Level-1 Tuning Problem (Static)(Calibrating Parameter Values)
Setting the length of Tabu tenure:– By Guessing ??– By Trial and Error ??– By using past experience as a rough guide ??
StaticLevel-3
TuningSearch Strategies
Level-2
ChoosingBest Configuration
Level-1
CalibratingParameter Values
DynamicTuning Problems
Tuning Tabu Search
Level-2 Tuning Problem (Static)(Choosing the best Configuration)
Choices of Local Neighborhood:– 2-opt ??– 3-opt ??– Very Large Scale Neighborhood (VLSN) ??
Choices of Tabu List:– Tabu moves ??– Tabu attributes ??– Tabu solutions ??
StaticLevel-3
TuningSearch Strategies
Level-2
ChoosingBest Configuration
Level-1
CalibratingParameter Values
DynamicTuning Problems
Tuning Tabu Search
Level-3 Tuning Problem (Dynamic)(Tuning Search Strategies)
Choices of Search Strategies:– Intensification ??– Diversification ??– Hybridization ??
Example:– Reactive Tabu Search (Battiti & Tecchiolli, 1994)
When and How to apply these strategies ??
StaticLevel-3
TuningSearch Strategies
Level-2
ChoosingBest Configuration
Level-1
CalibratingParameter Values
DynamicTuning Problems
Bottleneck !!!
Conventional Solution
Implement the metaheuristic
Evaluate its performance
Good or Give up
Not good
Stop
Modify the metaheuristic using past knowledge, past experiences, plus some instinct blindly.
(“Blind trial-and-error”)
Tuning: bottleneck in rapid development process (Adenso-Diaz & Laguna, 2005)
Automated Tuning Methods
Tool to automatically and systematically search for the best:Set of parameter values (level-1)Configuration (level-2)
Pros:Relieves the burden of tuning from human.
Cons:Treat metaheuristic as a black box.– Does not provide room for innovations…
Difficult to address level-3 Tuning Problem (for Dynamic Metaheuristic)Probably slow– if the number of possible configurations is high.
Examples:CALIBRA (Adenso-Diaz & Laguna, 2005)F-Race (Birattari, 2004)
Non-Automated Tuning Methods
Tool which allow human to diagnose the metaheuristicPros:
Make level-3 Tuning Problem for Dynamic Metaheuristic easierProvide room for innovations…
Cons:Human still need to do the job…Inconsistent results
Examples:Statistical Analysis, e.g.:– Fitness Landscape Analysis (Fonlupt et al, 1997)– Fitness Distance Correlation Analysis (Merz, 2000)
Human-Guided Tabu Search (Klau et al, 2002)Visualization of Search (Kadluczka et al, 2004)V-MDF (This work)
Visual Diagnosis TuningThe methodology for solving Tuning Problem
V-MDFThe tool to support Visual Diagnosis Tuning
Visual Diagnosis Tuning
Idea: Combine human intelligence and computer to produce good search strategies quickly.
Basic methodology of Visual Diagnosis Tuning:– {Cause-Action-Outcome} tuple:
• Diagnose incidents in search trajectory. (Cause)• Steer the search if necessary. (Action)• Instantly observe the impact of the his action. (Outcome)
– Example:• {passive searching – greedy random restart – arrive in good region}
– Possibly an effective strategy
• {solution cycling – decrease tabu tenure – solution cycling}
– Possibly an ineffective strategy
V-MDF: Distance Radar
Diagnose incidents in search trajectory– Visualizing search trajectory is difficult!
• Search space is large!
A special generic visualizer is needed: Distance Radar– Using the concept of “distance”
• Example:
Distance between two binary encoded solutions:hamming distance.
A = 110010B = 100011Distance = 2 bit flips.
V-MDF: Distance Radar
Main ideas of Distance Radar:– Record elite solutions along the search trajectory.
• Distance w.r.t Current solution• Recency w.r.t Current iteration• Objective Value w.r.t Current best objective value
– Current solution current position.– Elite solutions (Local Optimal) anchor points.– Approximate Tabu Search trajectory with these information.
V-MDF: Radar A
Distance Radarconsists of
Radar A and B.
This is Radar A. X-axis: Local Optimal
Y-axis: Distance
Plot distance and recency of these
local optimal against current solution and current interation
Distance Information
in Logarithmic
scale
This is a Recency Graph
to augment Radar A
X-axis: Local OptimalY-axis: Recency
Current solution is close to these elite solutions and they are recent.
Interpretation: exploring good regionIn Radar A, elite solutions are
sorted by Objective Value
V-MDF: Radar B
This is Radar B.It portrays distance
information from different angle. X-axis: Local Optimal
Y-axis: Distance
Plot distance and obj value of these
local optimal against current solution and
best so far
This is an Objective Value
Graph to augment Radar B
X-axis: Local OptimalY-axis: Objective Value
In Radar B, elite solutions are sorted by Recency
No cycling, objective value fluctuates. Interpretation: Tabu Search is
working correctly at the moment.
V-MDF: Distance Radar
Radar A, B, Recency and Objective Value Graph can be used together to draw more
information about the search trajectory
V-MDF: Remedial Actions
Series of non-improving moves observed…and it requires remedial action
For intensification, this is one of the correct trajectory
For Diversification, this is one of the correct trajectory
Rules selection phase
Visual Diagnosis Tuning Phase
V-MDF: Overall Architecture
Implement the metaheuristic in MDF framework (Lau et al, 2004)See also TSF in Metaheuristics: Progress as Real Problem Solvers.
Diagnose the metaheuristic against training instances using Distance Radar
Automatic extraction of Good Rules from Knowledge Baseto form the final metaheuristic algorithm
Add Rules to Knowledge Base
Apply the metaheuristic with good rules to whole test instances
Experiment using V-MDF
Task:– Tune a Tabu Search implementation for solving an
NP-hard Military Transport Planning (MTP) problem.
Knowledge base of rules after training.
Poor rules are discarded…Good rules form the final metaheuristic algorithm
Experiment using V-MDF
Objective Value versus Iteration for T4
0
100
200
300
400
500
1 51 101 151 201 251 301 351
Iteration
Ob
ject
ive
Val
ue
Before TuningObjective Value versus Iteration for T4
0
100
200
300
400
500
1 51 101 151 201 251 301 351
Iteration
Ob
ject
ive
Val
ue
Before Tuning
IntermediateObjective Value versus Iteration for T4
0
100
200
300
400
500
1 51 101 151 201 251 301 351
Iteration
Ob
ject
ive
Val
ue
Before TuningIntermediateAfter Tuning
The results of a training instance (minimizing problem)
Experiment using V-MDF
Tabu Search results
Summary
The Tuning Problem. (The Problem)– Taxonomy of tuning problems:
• Static vs Dynamic & 3 levels of complexity.– Current tuning methods:
• Automated vs Non-automated
Visual Diagnosis Tuning (The Methodology)– {Cause-Action-Outcome}.
Visualizer for MDF (V-MDF) (The Tool)– Distance Radar and its usage.– Overview of V-MDF– Generic (not restricted to one problem).– Useful especially for new problems.