17. Spawning Information
Agents on the Web
전기공학부
96420-113 박정혁
문제 제기
 1. 처리능력 밖의 작업이 Agent에게 주어졌을 때, 또는
능력 안의 일이지만 Overload 되었을 때.
– Task delegation to other agents
– delegation to similar agents
But, what if all the others are overloaded as well ???
 2. 필요로 하는 정보가 Remote host 에 존재하고 양이
많을 때, 어떻게 통신부담을 줄일 것인가?
–
대안 1
 Predictive Agent Cloning
– simple to create
 단점
– 원래 Agent와 같은 작업만 수행
– 복잡도로 인한 수행능력 저하
대안 2
 1. Single overload => pass to others
 2. All agents overload => create an agent,
migrate to other host
 3. Capability missing => create new agent
 4. Large remote information
=> migrate or Spawn !!!
기반 : RETSINA
 Providers
 Requesters
 Middle-agents
providers
middle-agents
requesters
Why is spawning necessary ?
 Cloning 기능과 더불어 data전송 최소화,
불필요한 공간의 절약 => 효율 향상
 Local spawning
– agent overload = machine overload
– idle processing capacity 활용
Spawning 조건
 Self-awareness
– on internal level
• 자료의 사용여부와 사용 part
• 내부 상태 (e.g. processor 활용도, 시간…)
– on MAS level
• middle agents 통해 원거리의 자료 파악
Spawning Procedure
 Reasoning before spawning
 Task splitting
 Spawning
 Reasoning after spawning
CPU, memory and communication loads
Simulation : The Cloning Case
When to Spawn ?
 Remote data : Ir
 Information produced by task : Ip
 Size of spawn agent : S
 raw data vs. local processing data : α(t)
Conclusion
 Locally or remotely create spawn agents
 A copy or a specifically tailored agent
 overload or large remote information
 For large numbers of tasks, spawning
significantly enhances the performance of
MAS