Supplementary Material for: Sparsely Changing Latent States for Prediction and Planning in Partially Observable Domains A Relation to other RNNs

In Sec. 3 we set out to create an RNN LSTMs use two gates, i.e. a forget and an input gate, with the sigmoid activation function Nonetheless, it is not straight forward to apply our approach, outlined in Sec. 3, to GRUs and LSTMs. Our loss (see Eq. 5) punishes non-zero gate activation. Thus, their gating function would need to be modified or replaced, e.g. by our ReTanh gate GateL0RD attempts to overcome the outlined downsides of using LSTMs and GRUs with our proposed latent state regularization. B.2 - B.5 provide further details specific to each Billiard Ball scheduled sampling (Sec. The best learning rates for all experiments are listed in Table 1.