Measuring Information Transfer in Online Adaptation Process of Recurrent Neural Networks

In this paper, we propose a simple model that focuses on the adaptation process of an agent to an unknown system in an online manner. The agent is equipped with a recurrent neural network, and by controlling the dynamics of the interacting system, it should predict its state in one-step prediction. To quantitatively characterize the interaction modality between the agent and the interacting system, we used transfer entropy. As a result, by varying the nonlinear parameter of the interacting system and the coupling strength, we numerically show that the adaptation dynamics can be distinguished between an agent-driven and a non-agent-driven dynamics.