A Convolution and Attention Based Encoder for Reinforcement Learning under Partial Observability