A kinematic model with a self‐learning capability of a series of motions based on instinctive incentive
Haruo Yoda, Takafumi Miyatake, Hitoshi Matsushima · Systems and Computers in Japan · 1991
Abstract The process of self‐learning of motions by animals has been modeled by a three‐layered neural net, and its function has been verified. In this model the neural net takes the place of a controller which receives the current shape and posture of an animal as input and feeds out the next optimum motion as output. To give a self‐learning capability to this neural net, this model includes a function to add a random number to the output of the neural net, an ability to evaluate the change of its shape and posture according to its instinct (evaluation function) and to generate appropriate training patterns. When a learning takes place following this model, an optimal motion sequence based on the instinct is formed automatically in the weights of the neural net. By computer simulation it has been observed that a starfish learns the sequences of getting‐up motions successfully and becomes skillful as it gains experience.