| Presentation | 2012-06-28 Learning Motor-visual Dynamics and Solving Linearlized Bellman Equation for Robot Control Ken KINJO, Eiji UCHIBE, Junichiro YOSHIMOTO, Kenji DOYA, |
|---|---|
| PDF Download Page |  PDF download Page Link |
| Abstract(in Japanese) | (See Japanese page) |
| Abstract(in English) | Recently, Todorov [1] proposed a technique to strictly linearize a Bellman equation under a instruction on the cost function by exponential transformation of the variable. This enables deriving the value function and the optimal control law analytically, because the Bellman equation became an eigenvalue problem. In continuous state space case, a linearized bellman equation is required to solve an eigenfunction problem, Todorov has already shown a technique for deriving the eigenfunction by using the functional approximation [2]. Although these techniques are attractive for application to real system like robot, They assume that the dynamics of the system is already-known. In a real system, it is rare that they are already-known. It investigate only low dimensionality like swing-up balancing task. In this paper, We proposes a method for deriving an optimal control law from the estimated motor-visual dynamics from the sequence of experienced states and action and apply this method to real system with high state-actions space. In a visual guide task, Robot learn appropriate behavior and obtain better controller than LQR when the problem setting is equivalent to LQR. |
| Keyword(in Japanese) | (See Japanese page) |
| Keyword(in English) | desirability function / system identification / optimal control |
| Paper # | NC2012-4 |
| Date of Issue |
| Conference Information | |
| Committee | NC |
|---|---|
| Conference Date | 2012/6/21(1days) |
| Place (in Japanese) | (See Japanese page) |
| Place (in English) | |
| Topics (in Japanese) | (See Japanese page) |
| Topics (in English) | |
| Chair | |
| Vice Chair | |
| Secretary | |
| Assistant | |
| Paper Information | |
| Registration To | Neurocomputing (NC) |
|---|---|
| Language | JPN |
| Title (in Japanese) | (See Japanese page) |
| Sub Title (in Japanese) | (See Japanese page) |
| Title (in English) | Learning Motor-visual Dynamics and Solving Linearlized Bellman Equation for Robot Control |
| Sub Title (in English) | |
| Keyword(1) | desirability function |
| Keyword(2) | system identification |
| Keyword(3) | optimal control |
| 1st Author's Name | Ken KINJO |
| 1st Author's Affiliation | Graduate school of Informatin, Nara Institute of Science and Technology:Neural Computation Unit, Okinawa Institute of Science and Technology() |
| 2nd Author's Name | Eiji UCHIBE |
| 2nd Author's Affiliation | Neural Computation Unit, Okinawa Institute of Science and Technology |
| 3rd Author's Name | Junichiro YOSHIMOTO |
| 3rd Author's Affiliation | Neural Computation Unit, Okinawa Institute of Science and Technology:Graduate school of Informatin, Nara Institute of Science and Technology |
| 4th Author's Name | Kenji DOYA |
| 4th Author's Affiliation | Neural Computation Unit, Okinawa Institute of Science and Technology:Graduate school of Informatin, Nara Institute of Science and Technology |
| Date | 2012-06-28 |
| Paper # | NC2012-4 |
| Volume (vol) | vol.112 |
| Number (no) | 108 |
| Page | pp.pp.- |
| #Pages | 6 |
| Date of Issue | |