Frozen Recurrent Networks Support Long-Term Memory in POMDPs

Experimental results show that frozen recurrent networks, inspired by reservoir computing, provide sufficient memory for solving partially observable Markov decision processes. These fixed modules achieve competitive control performance and support recall over hundreds of steps without joint training of the recurrent weights.