_neural_evolutionary_agent not learning

[HermesMG]

Hello:
I am executing the "_neural_evolutionary_agent,ipynb" notebook _first chapter- using 10 epochs and many times the agent does not learn, that is the reward is 0 for each step, and does not change.
In a few ocassions the reward changes with the epoch.
Can anybody explains me why I observe such strange results?
Thank you in advance
Hermes

[praveen-palanisamy]

Hi @HermesMG ,
For the neural evolutionary agent, 10 epochs is a very low number to train for convergence. In general for (approximate) Cross Entropy Methods especially with a neural network, the number of epochs needed is quite larger depending on the problem and the neural network's parameter count.
It is common to run for 200k+ epochs but if you are training with the default Brain and Agent definition and for the Gridworld-v0 environment, you might see convergence with ~100k epochs or less.
If training for that long is not feasible on your end, please try with a minimum of 1000-5000 epochs to see enough variance in the rewards.