| Project/Area Number |
22860041
|
|---|
| Research Category |
Grant-in-Aid for Research Activity Start-up
|
| Allocation Type | Single-year Grants |
|---|
| Research Field |
Control engineering
|
|---|
| Research Institution | Hiroshima University |
|---|
Principal Investigator |
SATOH Satoshi 広島大学, 大学院・工学研究院, 助教 (60533643)
|
|---|
| Project Period (FY) |
2010 – 2011
|
| Project Status |
Completed (Fiscal Year 2011)
|
| Budget Amount *help |
¥3,146,000 (Direct Cost: ¥2,420,000、Indirect Cost: ¥726,000)
Fiscal Year 2011: ¥1,508,000 (Direct Cost: ¥1,160,000、Indirect Cost: ¥348,000)
Fiscal Year 2010: ¥1,638,000 (Direct Cost: ¥1,260,000、Indirect Cost: ¥378,000)
|
|---|
| Keywords | 非線形制御 / 確率制御 / 学習制御 / 歩行ロボット / 確率システム |
|---|
| Research Abstract |
Firstly, we have extended conventional deterministic Hamiltonian systems to stochastic ones, and have clarified some of their properties such as stochastic passivity and symmetry. This symmetry is a property that a state-space realization of the variational system of a stochastic Hamiltonian system coincides with a time-reversal version of that of the adjoint one. Although the adjoint system is described by a backward stochastic differential equation, and it is generally difficult to be solved, the symmetric property enables us to solve the problem by calculating a corresponding forward stochastic differential equation. Secondly, we have proposed a new learning optimal control method based on the symmetric property of stochastic dynamical systems. This method generates an optimal feedforward input and a corresponding optimal trajectory, which minimize(at least locally) the expectation of a given cost function by iteratively updating the input in the gradient direction of the expectation of the cost function. In calculating the gradient, the adjoint system has to be calculated, and it generally requires the precise knowledge of the plant system. However, the proposed method can solve the problem without information of the plant system by utilizing the symmetric property.
|
