multi-objective optimization in rule-based design space exploration (ase 2014)
DESCRIPTION
Design space exploration (DSE) aims to find optimal design candidates of a domain with respect to different objectives where design candidates are constrained by complex structural and numerical restrictions. Rule-based DSE aims to find such candidates that are reachable from an initial model by applying a sequence of exploration rules. Solving a rule-based DSE problem is a difficult challenge due to the inherently dynamic nature of the problem. In the current paper, we propose to integrate multi-objective optimization techniques by using Non-dominated Sorting Genetic Algorithms (NSGA) to drive rule-based design space exploration. For this purpose, finite populations of the most promising design candidates are maintained wrt. different optimization criteria. In our context, individuals of a generation are defined as a sequence of rule applications leading from an initial model to a candidate model. Populations evolve by mutation and crossover operations which manipulate (change, extend or combine) rule execution sequences to yield new individuals. Our multi-objective optimization approach for rule-based DSE is domain independent and it is automated by tooling built on the Eclipse framework. The main added value is to seamlessly lift multi-objective optimization techniques to the exploration process preserving both domain independence and a high-level of abstraction. Design candidates will still be represented as models and the evolution of these models as rule execution sequences. Constraints are captured by model queries while objectives can be derived both from models or rule applications.TRANSCRIPT
Multi-Objective Optimization inRule-Based Design Space
Exploration
Hani Abdeen, Dániel Varró, Houari Sahraoui, András Szabolcs Nagy, Csaba Debreceni, Ábel Hegedüs, Ákos Horváth
International Conference on Automated Software Engineering (ASE 2014) Västerås, Sweden, September 15 - 19, 2014
Design Space Exploration (DSE)
• Special state space exploration– Potentially infinite state space– cannot put upper bound on the number of model elements used in a design
candidate (elements are created and deleted during exploration).
Design Space Exploration
Design Alternative 1
Design Alternative 2
Design Alternative 3
Design Alternative 4
Goals
Global Constraints
Operations
Initial Design
2
Rule-based Design Space Exploration
• Objectives : complex model metrics calculated by model queries• Cost calculations may depend on the seq. of transf. rules• Multiple objectives
Design Space Exploration
Seq. of Transf. Rules 1
Seq. of Transf. Rules 2
Seq. of Transf. Rules 3
Seq. of Transf. Rules 4
Model queries as Goals
Model queries as Constraints
Transf. rules as Operations
Initial Model as a graph
Modified model
Operation
Initial model
Solution model
Constraints violated
Goals satisfied
3
Motivating example: Smart Building
• Reconfiguration of supervising cyber-physical systems (CPS) – Offices to rent with highly
configurable services – Services to deploy on both
embedded and virtual computational units
– Requests may change over time– Certain faulty devices may no
longer function
4
Motivating example: Smart Building
Architecture5
Smart Building: configuration modelServices and Requests
6
(a) Services
(b) Two examples on company requests
6
Approach : Basic Conceptsdom
ain
Model
New/Modified requirements
candidate Models
Incorporate changes
Exploration rules
Search ideal candidate
Objectives
Constraints
7
Approach : Basic Concepts
MnMM nrere ...10Rn1
Rn
8
Approach : Basic Concepts
• Constraints– Graph patterns to search for with model queries– For smart buildings
• Constraints define valid or invalid configurations
Positive Positive Positive Negative
9
Approach : Basic Concepts
Positive for well-formedness constraintsNegative for ill-formedness constraints
10
Approach : Basic Concepts
• Objectives– Satisfying all constraints is a top-level
objective – Depend on the domain – Derived from
• Models (Model objectives) • Rule applications (Trajectory objectives)
11
Approach : Basic Concepts
12
Approach : Basic Concepts
13
Approach : Algorithm• NSGA-II
14
Approach : Algorithm
• Constraint-handling strategy– A solution s1 constrained-dominate a solution
s2, if:1. s1 is valid and s2 is not 2. both s1 and s2 are invalid, but s1 has a smaller
overall constraint violation 3. both s1 and s2 are valid and s1 dominates s2 with
the usual domination function wrt. the optimization objectives.
15
Approach : Algorithm
• Constraint-handling strategy
Rank 1Rank 2
…Rank 3
…Rank N
Rank 4
Rank 1Rank 2
…Rank 3
…Rank N
Rank 4
1Same fulfillment
value2
Other optimizationobjectives
Constraint fulfillment
16
Approach: Algorithm
• Genetic operators
One-point crossover Cut-and-splice crossover Permutation crossover
Add mutation Delete mutation Swap mutation
17
Experimental Evaluation• Data
– Smart building (re)configuration
• Comparison with: 1. Random simulation (Random)2. Fixed priority local search (FPLS) strategy
• Requests – Increasing number of rooms (4, 6, 8, 12)
Problem size (rooms) Model size (graph elements)
4 130
6 200
8 230
12 330
18
Experimental Evaluation
• Initial model– Requests with requirements, and application
and host types – No instances
• DSE process 1. create a sufficient number of application and
host instances 2. allocate application instances to host instances3. start and stop the application instances
19
Experimental Evaluation
20
Experimental Evaluation Results
Quality of NSGA solutions produced in 30 runs for different problem sizes, as measured by normalized constraints’ fulfillment, cost and computer server utilization
21
Experimental Evaluation Results
• Higher constraint fulfillment with NSGA• Lower Cost with NSGA• Results statistically significant (two-tailed Wilcoxon
tests)
0.3
0.4
0.5
0.6
0.7
0.8
0.9
1
201 286 315 377 383 550
Norm. Constraint Fulfillment
Cost
Constraint Fulfillement vs. Cost
NSGA
FPLS
Random
22
Experimental Evaluation Results
• Higher computer server utilization with NSGA• Results statistically significant (two-tailed Wilcoxon
tests)
23
00.10.20.30.40.50.60.70.80.9
1
201 286 315 377 383 550
Computer Server Utilization
Cost
Computer Server Utilization vs. Cost
NSGA
FPLS
Random
ConclusionDomain Model Requirements Exploration Rules
New Requirements
M0
Graph Patterns
Constraints
Optimization Objectives
On Model
On Trajectory
client
Domainexperts
………
Pareto solutions
…
Fittest Models
…
…
NSGA-based evolution withConstraint-handling strategyNSGA-based evolution withConstraint-handling strategy
Initial population
24