Click here to load reader
View
0
Download
0
Embed Size (px)
Computational Intelligence For Optimization
Deepak SharmaDeepak Sharma Research Fellow
Department of Electrical and Computer Engineering National University of Singapore
Email: eledeepa@nus.edu.sg
Overview of TalkOverview of Talk Genetic Algorithm for Designing Introduction and Procedure of GA Introduction and Procedure of GA Application in designing mechanism Approach Results and Discussion Conclusions
M l i A M d li f h d li Multi-Agent Modeling for scheduling Introduction of MAS and Thermal generation scheduling Problem formulation Problem formulation Approach Results and Discussion Conclusions
2 SIS@SMU
SIS@SMU3
Borrowed from the lecture notes of Dr Dipti Srinivasan @ NUS
Genetic Algorithm for DesigningGenetic Algorithm for Designing Introduction Computerized search and optimization technique Computerized search and optimization technique Based on Darwin’s principal of natural selection and genetics
inheritance Works on the population Does not require gradient information Can avoid premature convergence to local optima's Can avoid premature convergence to local optima s Suitable for complex problems
Example: Mixed variables, continuous or discontinuous search space, non-linearity, multiple objectives etc.
However, requires more computation time than point-by-point search methods
SIS@SMU4
Procedure Initialize Procedure Initialize Population
Fitness evaluation of individuals
Population Based Meta-Heuristic Algorithm Initial random population of candidate
solutions
Ranking
Termination
Parent Population
so ut o s Fitness evaluation of individuals, e.g., objective
and constraint violation calculations Ranking: Assign rank to every individual in the
population based on fitness values Termination criterion
met?
SelectionElimination
No
Yes
ppopulation based on fitness values Selection: Survival of fittest by choosing good
solutions based on ranking Crossover: Create new individuals from the
selected parents Terminate
selected parents Mutation: Perturb the newly created individuals Fitness evaluation of new population Ranking of new population Report
results CrossoverRanking Elimination to propagate good individuals
Termination criterion Report results
results
SIS@SMU5
MutationFitness evaluation
Child Population
Application in DesigningApplication in Designing Goal: To design a mechanism Path generating compliant mechanismg g p
One piece flexible elastic mechanism that has to generate user defined path
Identify the locations where y material has to be filled optimally
Advantage over traditional mechanism Joint-less and monolithic structureJ Less friction wear and noise Light weight Easy to manufacture without assemblyasy to a u actu e w t out asse b y
Applications Micro-actuators Micro switches
SIS@SMU6
Micro-switches
Designing criterion in literature Minimizing the gap between actual and user
defined paths (single or multi objective)defined paths (single or multi-objective) Limitation: Without limiting the gap, actual
and prescribed paths can be apart from each other
Motivation Functionality of mechanism should be represented in terms of
constraints so that every feasible solution can be referred as path generating compliant mechanism
Can solve the problem for optimal distribution of material in the design domain
P d A h Proposed Approach Objective:
Minimum weight Constraints:
SIS@SMU7
Constraints:
Characteristics of given designing problemg g g p •Discrete search space
•Mixed-variables •Non-linear
•C t i d ti i ti bl•Constrained optimization problem
SIS@SMU8
Multi-ObjectivizationMulti Objectivization Definition: Provide solution to single objective optimization by adding helper g j p y g p
objective(s) and deal with the problem as multi-objective optimization
Advantages G id h h f l i l i h id l l i Guide the search for evolutionary algorithms to avoid local optima
Maintain the diversity in the population
Further Benefit Multiple optimal solutions
Bi-Objectives: Minimize weight of structure (Primary objective) Minimize supplied input energy to structure (Secondary Objective)
SIS@SMU9
Customized Genetic AlgorithmCustomized Genetic Algorithm Structure
representationrepresentation
Domain specific initial population
SIS@SMU10
Contd.Contd. Chromosome
Shape C diti
Connectivity and repairing techniques To connect topology Conditions To connect topology Single point connectivity Floating element
2D crossover operator
Finite element analysis S d d fl i l i Stress and deflection analysis
Parallel Computing Master-Slave architecture F d i th t ti
SIS@SMU11
For reducing the computation time
Contd.Contd. Local search: Post processing Mutation and hill-climbing Mutation and hill climbing Weighted sum method to reduce the bi-objective problem into
single objective
Weights Weights
SIS@SMU12
Results and DiscussionResults and Discussion Comparison Customized Initialization of structure shapes Customized Initialization of structure shapes Random Initialization of structure shapes However, rest of the customized GA is same for both cases
SIS@SMU13
Contd.Contd. Performance test R-indicator: Closeness to R indicator: Closeness to
the reference Pareto-Optimal front H l i di S d Hypervolume-indicator: Spread and closeness
Path traced by solutions
SIS@SMU14
Contd.Contd. Shapes of structures
SIS@SMU15
Contd.Contd. Optimum conditions Main Findings
Formulation that can guarantee feasible designs
Multiple solutions for decision making
Computation time
ec s o a g Optimum set of conditions Better performance of
customized GA against customized GA against random initialization based GA
Customized schemes able Customized schemes able to generate practically feasible designs
P ll l t i d GA
SIS@SMU16
Parallel customized GA
Multi-Agent Modeling for SchedulingMulti Agent Modeling for Scheduling Multi-Agent System (MAS)
System of flexible autonomous agents where these agents are loosely connected and act in the environment to achieve their goals
Broadly classified approaches for exploiting MAS Building flexible-robust-extensible system Modeling the large and complex system S b fi Some benefits Increase computation speed because of parallel computation and asynchronous
operations Graceful degradation of system when one or more agents failg y g Scalability and flexibility of adding and removing agents Reusability of agents
Critical challenges U bl b h i i h i i Unstable behavior in changing environment
Converge to sub-optimal solution due to limited visibility Action or decision taken by an agent may not be suitable for another agent
which can be reduced by sharing information. But the problem is when to
SIS@SMU17
y g p communicate and to which agent
Difficult in debugging and testing parallel and distributed system
Day Ahead Thermal Generation SchedulingDay Ahead Thermal Generation Scheduling In restructured power system Generator companies (GENCOs) are independent and Generator companies (GENCOs) are independent and
autonomous These companies compete with each other for selling the power and
reserve in the power market to earn profitreserve in the power market to earn profit Aim: To maximize their profit by generating desired amount of power
and reserve by optimally scheduling the thermal generators I bl f GENCO h l b dd k It is important problem for GENCO while bidding in power market
GENCO performs a day ahead scheduling task based on the forecast data We assume that the forecast data is given
SIS@SMU18
Contd.Contd. Scheduling involves two decision making processes ON/OFF status of every generator (Unit Commitment) ON/OFF status of every generator (Unit Commitment)
In peak load hours, more number of units should be committed that can earn profit
I ff k l d h i t it 1500
2000
( M
W )
In off-peak load hours, appropriate units should be committed that can earn profit
Because of the physical constraints, any t t t t h t d
0
500
1000
0 4 8 12 16 20 24Lo ad
D em
an d
generator can not start or shut-down, immediately.
Thus