Download 3. The Approach

Survey
yes no Was this document useful for you?
   Thank you for your participation!

* Your assessment is very important for improving the workof artificial intelligence, which forms the content of this project

Document related concepts

Genetic testing wikipedia , lookup

Gene expression programming wikipedia , lookup

Adaptive evolution in the human genome wikipedia , lookup

Dual inheritance theory wikipedia , lookup

Microevolution wikipedia , lookup

Population genetics wikipedia , lookup

Koinophilia wikipedia , lookup

Transcript
DOSHISHA
UNIVERSITY
XML-based
Genetic Programming Framework:
Design Philosophy, Implementation
and Applications
24 May 2017
1
DOSHISHA
UNIVERSITY
Outline
1.Introduction
2. Objective
3. Proposed approach
4. Verification results
5. Applications
6. Conclusion
24 May 2017
2
1. Introduction: the Problem
DOSHISHA
UNIVERSITY
A) Promptly developed software models of the evolved artifac
B) Fast running offline (phylogenetic) learning
via simulated evolution (e.g. GP)
The Needs
24 May 2017
3
1. Introduction: the Problem
DOSHISHA
UNIVERSITY
A) Promptly developed software models of the evolved artifacts
B) Fast running offline (phylogenetic) learning
via simulated evolution (e.g. GP)
The Needs
Discrepancy, Gap
The Reality
A) Slow development time of evolutionary systems
(specific semantics)
B) Notoriously poor performance of GP
(populations, generations, independent runs)
24 May 2017
4
2. The Objective
DOSHISHA
UNIVERSITY
A) Promptly developed software agents
B) Fast running offline (phylogenetic) learning
via simulated evolution (e.g. GP)
The Needs
Discrepancy, Gap
The Reality
The Reality
A) Quicker
development time GP
B)Slow
Better
performance
of GP
A)
development
time ofcharacteristics
evolutionary systems
(specific semantics)
B) Notoriously poor performance of GP
(populations, generations, independent runs)
24 May 2017
5
3. The Approach
DOSHISHA
UNIVERSITY
Quicker development time of GP ?
History of “Reuse of Software Blocks” in
Software Engineering:
• loops,
• procedures, functions (incl. recursions),
• modules (units),
• objects,
• component objects
Component objects (CO):
• appears to be an object of the IDE which
incorporates them,
• binary standard (language-independent)
24 May 2017
6
3. The Approach
DOSHISHA
UNIVERSITY
Focusing on representation of genetic programs:
A) Standard DOM-parsing tree
and XML text.
B) CO: DOM-parser with
built-in API for dealing with
genetic programs.
24 May 2017
7
3. The Approach
DOSHISHA
UNIVERSITY
Advantages
A) Significant reduction of the time consumption
of software engineering of GP using build-in API
for creating and manipulating genetic programs.
24 May 2017
8
3. The Approach
DOSHISHA
UNIVERSITY
Issue:
How to represent the allowed syntax (i.e. to
reduce the search space) of GP?
• In the program source of GP-system
(modifications by expert, recompilation, etc…) ?
• As an external text with well-known format?
Employing XML facilitates the second choice.
24 May 2017
9
3. The Approach
DOSHISHA
UNIVERSITY
Advantages
B) Increase of efficiency of execution of XGP:
Reducing the computational effort as a result of
generic support for the idea of pruning the solution
space via strongly typed GP.
How:
XML-schema as a standard, generic way to represent
the syntax of XGP.
24 May 2017
10
3. The Approach
DOSHISHA
UNIVERSITY
• Relationship between tree
nodes in XGP,
• Data types associated with tree
nodes
Fragment of XML Schema
<xs:simpleType name="VAR_TSpeed">
<xs:restriction base="xs:string">
<xs:enumeration value=“Speed" />
</xs:restriction></xs:simpleType>
<xs:simpleType name="OPER_TSpeed">
<xs:restriction base="xs:string">
<xs:enumeration value="GE" />
<xs:enumeration value="LE" />
</xs:restriction></xs:simpleType>
<xs:simpleType name="CONST_TSpeed">
<xs:restriction base="xs:integer">
<xs:minInclusive value="0" />
<xs:maxInclusive value=“22" />
</xs:restriction></xs:simpleType>
24 May 2017
11
3. The Approach
DOSHISHA
UNIVERSITY
Advantages
B) Increase of efficiency of execution of XGP parallelism:
• Improving the computational performance:
XML representation of both the schema and the
genetic programs is a feasible format for
migration of agents in parallel, distributed
computer architectures.
In-memory tree structures of GP cannot be transferred
between computing units in parallel architectures.
24 May 2017
12
DOSHISHA
UNIVERSITY
3. The Approach
Memory Structure (DOM)
Text (XML)
Straightforward
Mapping
24 May 2017
13
3. The Approach
DOSHISHA
UNIVERSITY
Structure of XGP-framework
GP Manager (selection,
Implications:
crossover,
and mutation)
Simulation Boards
Reuse of GP(only
Manager across the applications,
Domain•Independent
(evaluation)
• Parallel
Simulation
Boards
XML Schema
need
to be
Domain-specific
updated)
24 May 2017
14
3. The Approach
DOSHISHA
UNIVERSITY
Example – Evolution of Behavior of Agents in MAS
Parallel Implementation via Boss-Workers Model
Genetic program
(XML)
Fitness
GP Manager
(selection, crossover,
and mutation)
Simulation Boards
(evaluation)
24 May 2017
15
4. Verification Results
DOSHISHA
UNIVERSITY
• Development time for the initial prototype of
XGP (from scratch): several [person*days]
24 May 2017
16
4. Verification Results
DOSHISHA
UNIVERSITY
• Porting time (employing XGP for already
developed simulation board): less than one hour
XML Schema
File
24 May 2017
17
DOSHISHA
UNIVERSITY
4. Verification Results
• Computational Effort of XGP: Reducing the
search Space (XML Schema)
Probability of Success for Evolution of XGP with
(STGP) and without (LP, LPA) strong types
1.0
STGP
LP
LPA
p(t)
0.8
0.6
0.4
0.2
0.0
0
24 May 2017
8000
16000 24000 32000
Indiv iduals ev aluated
40000
18
DOSHISHA
UNIVERSITY
5. Applications
Evolution of Agents Behavior in MAS
GP Manager
Domain Neutral
24 May 2017
MAS Simulation
Board
Domain Specific
19
DOSHISHA
UNIVERSITY
5. Applications
Evolution of Agents Behavior in MAS
XML
representation
of GP
24 May 2017
20
5. Applications
DOSHISHA
UNIVERSITY
Evolution of Locomotion of Snakebot
GP Manager
Domain Neutral
DOM
representation
of GP
Simulation Board
Domain Specific
24 May 2017
21
DOSHISHA
UNIVERSITY
5. Applications
Evolution of Neural Networks
GP Manager
Domain
Neutral
Simulation Board
Domain Specific
24 May 2017
XML
representation
of GP
22
5. Applications
DOSHISHA
UNIVERSITY
Evolution of Driving Agent
Camera
(perceptions
of the agent)
PC
(driving agent)
Control
Loop, 100ms
Car
(1/24 Scale Model)
24 May 2017
Remote Control
(agent’s actions)
23
DOSHISHA
UNIVERSITY
5. Applications
Evolution of Driving Agent
GP Manager
Domain
Neutral
DOM
representation
24 May 2017 of GP
Simulation Board
Domain Specific
24
5. Applications
DOSHISHA
UNIVERSITY
Interactive Evolution of Postures of Aibo Robot
GP Manager
Domain
Neutral
DOM
representation
of GP
24 May 2017
Simulation Board
Domain Specific
25
DOSHISHA
UNIVERSITY
5. Applications
Interactive Evolution of Room Colors
GP Manager
Domain
Neutral
DOM
representation
24 May 2017 of GP
Simulation Board
Domain Specific
26
DOSHISHA
UNIVERSITY
5. Applications
Evolution of Human-Relation Networks
GP Manager
Domain Neutral
24 May 2017
Simulation Board
Domain Specific
27
6. Conclusion
DOSHISHA
UNIVERSITY
Proposed DOM/XML-Based Portable Genetic
Representation in XGP
A)Reduced Development Time
• Managing genetic program via standard DOM
parsers with built-in API
B) Easy Porting to New Applications
• Reusing the very General, Domain-Independent
GP Manager,
• Modifying the XML-schema only.
24 May 2017
28
6. Conclusion
DOSHISHA
UNIVERSITY
Proposed DOM/XML-Based Portable Genetic
Representation in XGP
C) Improved Execution Time of XGP
• Reducing Computational Effort: Limiting solution
space using strongly typed GP and offering
generic support via XML schema,
• Improving Computational Performance: Generic
support of distributed (web-compliant)
implementation of GP.
Drawbacks?
• Fitness evaluation – parsing of XML/DOM tree
and navigating among the nodes…
24 May 2017
29