Deep Q-Learning with Atari Games

I was trying to train SpaceInvadersNoFrameskip-v4 env locally but have an entry lvl gpu with less than great memory. I tried to solve it with a simple bash script that reloaded the previously trained agent after ~125k steps. The results seemed to be inline with it creating a new agent everytime. I used the -i attribute to load the previously trained agent. Why didn’t this work?