Same Algorithm Different Epsilon

As some already guessed it, A and B are actually the same RL algorithm (A2C), sharing the exact same code, same hardware, same hyperparameters... except the epsilon value to avoid division by zero in the optimizer (one is `eps=1e-5`, the other `eps=1e-7`)

A technical screenshot jokes that two supposedly different reinforcement-learning algorithms are identical in every meaningful way, differing only in the optimizer’s epsilon value. The punchline is that a tiny implementation detail is presented as the only distinction between A and B, humorously exaggerating how similar they are. EN: A technical screenshot jokes that two supposedly different reinforcement-learning algorithms are identical in every meaningful way, differing only in the optimizer’s epsilon value. The punchline is that a tiny implementation detail is presented as the only distinction between A and B, humorously exaggerating how similar they are.

• @ithueti / 3198 2020-11-19