|Budget Amount *help
¥2,200,000 (Direct Cost : ¥2,200,000)
Fiscal Year 2000 : ¥500,000 (Direct Cost : ¥500,000)
Fiscal Year 1999 : ¥700,000 (Direct Cost : ¥700,000)
Fiscal Year 1998 : ¥1,000,000 (Direct Cost : ¥1,000,000)
In modelling an autonomous decentralized production system, which consists of machine tools, transportation equipment, automated warehouses, and data processors, as a multiple agents system, AGV, one of transportation equipment, occupies an important location in the system. AGV in the production system has large freedom because it is movable transportation equipment. This research focuses on acquisition of autonomous behavior in order to transport work-pieces by learning and adapting its dynamic factory environment.
During the years 1998-2000, the following topics are studied.
(1) Development of scene acquisition method for AGV for use of path acquisition
(2) Development of path acquisition method for AGV by use of Q-learning
(3) Establishment of communication protocol between AGVs to understand mutual behavior
For the first problem, SDM is introduced and applied to acquire future scenes in the factory required for AGV to drive itself autonomously. For the second problem, Q-learning is intr
oduced. By combining SDM and Q-learning, AGV can acquire proper path between given two locations. The research object for the last problem is to acquire AGV collision avoidance knowledge for AGV.To avoid collision between AGVs, it becomes necessary to mutually understand other AGV's behavior - communication protocol. Numerical experiments show Q-learning can acquire proper communication protocols for collision avoidance.
For the last year 2001, the following topics are studied.
(1) AGV cooperative transportation by pushing a bar-shaped object
(2) Collision avoidance with moving object by learning
Q-learning is implemented to solve both problems. Numerical experiments in the case that three AGVs push the bar-shaped object are performed. Results show that Q-learning gives AGVs good cooperative behaviours on the first problem. Q-learning taking consider in a time series of situation are developed and implemented. Numerical simulation shows that the proposed method acquires the prediction knowledge to catch up the moving object. Less