2006 Fiscal Year Final Research Report Summary
Studies on Agent-Oriented Scheduling on A Production System based on Adaptive and Learning Strategies
| Project/Area Number |
16560105
|
|---|
| Research Category |
Grant-in-Aid for Scientific Research (C)
|
| Allocation Type | Single-year Grants |
|---|
| Section | 一般 |
|---|
| Research Field |
Production engineering/Processing studies
|
|---|
| Research Institution | HOKKAIDO UNIVERSITY (2006)
Asahikawa National College of Technology (2004-2005) |
|---|
Principal Investigator |
FURUKAWA Masashi Hokkaido Univ., Grad. School of 1ST., Prof., 大学院情報科学研究科, 教授 (70042091)
|
|---|
| Project Period (FY) |
2004 – 2006
|
| Keywords | agent / adaptive / learning / evolution / small-world / TSP / production system |
|---|
| Research Abstract |
In 2004, an agent model that can select an adaptive strategy under a given situation is theoretically established and its validity was verified by numerical simulations.
In 2005, a small-world graph was introduced to design a production system environment. In this research, automatically guided vehicles (AGVs) are regarded as agents and their driving lanes are regarded as the environments. As a result, it was seen that efficiently driving lanes for AGVs can be designed by use of the small-world graph.
In 2006, agents' structures were reconsidered. Conventional agents' structures are designed based on an artificial neural network (ANN) or Q-learning that adaptively allow the agents to acquire event-driven roles. A back propagation and recurrent ANNs are well known as conventional ANN structures. However, these structures have redundant synapses because there graph structures are nearly similar to a complete graph. On the other hand, there are creatures that have neural networks with much smaller number of synapses than conventional ANNs and recently the neural network was discovered which realizes small-world graph. This means a process of evolution for the creature's neural network removed the redundant synapses. Based on this fact, a new neural network structure for an adaptive and learning agent was designed, which realizes a small-world graph within itself. After embedding the new neural network structure into the agents, agents' performance was examined by numerical simulation. As a test-bed problem, a light pursuit problem was employed. By comparing the new structure with the back propagation and recurrent neural network structures, it is verified the new structure, based on a small-world graph, shows better performance than others in term of learning speed and that it consists of smaller synapses than others.
|
Research Products
(10 results)
