The behavioral rule we propose can be formalized in a straightforward manner in a completely general way as well. Denoting by uit the i-th component of a motor command of an agent at time t, the same component at time t+1 is given by
where R is a random variable such as a white gaussian noise, ηi means the intensity of the noise in the i-th component of the motor command, and ΔAt expresses how much the agent improved its state during the t-th time step.
We have confirmed and theoretically proven that the agent can follow a gradient of an evaluation function despite that the rule is extremely simple thanks to stochastic resonance.
The minimalistic behavioral rule is extremely simple and does not require almost all information about robot's hardwares and environments. Therefore, its property will be maintained even if the environment and/or the robot's hardware was drastically changed. In this research, we gave several severe hardware accidents of a mobile robot (such as blowouts, bend of axle shafts, and loss of control signals) and confirmed that the property of the minimalistic behavioral rule could be maintained.
To detect weak signals smaller than a sensor resolution has been often studied as engineering applications of stochastic resonance. To exploit stochastic resonace, the noise intensity should be adequate to the intensity of the weak signal, but the intensity of the signal is probably unknown in advance.
In this research, we propose the method to optimize the noise intensity by using spurious correlation among redundant sensors without knowing the input signal.
Stochastinc resonance can be very useful for control of distributed systems consisting of very simple elements. In this research, we focused on an actuator consisting of elements that can only fully contract and fully relax as the same as skeletal muscles in our body, and proposed a very simple system to continuously control the output force.
© Copyrights Shuhei Ikemoto. All Rights Reserved