At Robomech2017 in Fukushima, I presented a following content.
“Multi-Objective Switchable Reinforcement Learning by using Reservoir Computing”
In this presentation, I utilize a reservoir computing, which is a type of recurrent neural network without learning the weights of input and reservoir layers. I proposed a learning method to mitigate catastrophic forgetting by utilizing inherent dynamics of the reservoir computing: namely the inherent dynamics change behavior depending on inputs.