In general I'm very very happy with the new paradigm, feel like some kind of pioneer, because I've never seen this for agent training tasks! Now I'll try to formulate everything as energy-based model, maybe training will be much simpler! If you have ideas what approaches to try - write!
PS: This is why I rejected your stupid Reinforcement Learning at the very beginning and chose genetic algorithms, because can you imagine me conducting the same experiment with RL? I would never do it, I wouldn't have enough brains, time, or computer to train Energy-based model through RL. That's why they haven't earned such popularity, because they're damn hard to train. But for genetic algorithm I coded this in half an hour, and during training it couldn't care less. This is called FREEDOM.