Future Prediction with Hierarchical Episodic Memories under Deterministic and Stochastic Environments

In agreement with Bond’s suggestion, we consider that episodic memories are hierarchized autonomously by simple rule. In this research, our model solves maze tasks. Each episodic memory corresponds to the model’s each track. In our previous research, we suggested that our model concatenates episodic memories into one long episodic memory. Our previous model showed successful prediction of any long periodical and deterministic environmental changes with editing (selecting and concatenating with adequate timing) stored episodic memories autonomously. However, the previous models could not select adequate actions under a stochastic environment like POMDPs. Here, we suggest hierarchical episodic memories implement into the model. It is shown that the model improved not only their action under POMDPs but also prediction of long-term environmental change and incremental learning.