As I mentioned in the other post, the curiosity bots can help with this.
They're rewarded by exploring, not by values in game.
Maybe alone they're not enough, but in conjunction with other things, I bet we could beat Zelda. I had the bot exploring enough to find the first dungeon of LoZ.
Yeah, topology and weights. It's highly subject to initial conditions. You almost need another NEAT network to evolve the initial conditions. I believe it's turtles all the way down.
The article is from last year but it's still extremely valuable and interesting.
Exploring this topic is currently my primary hobby. Specifically, I've been using OpenAI's retro (Sonic, Contra, Mario, Donkey Kong and, more recently FZero) and comparing the ancient NEAT with more fashionable stuff like DQN, PPO, A3C and DDPG.
With my extremely limited experience, NEAT seems to outperform all of these other algorithms. I believe the advantage is the potential for strange/novel network structure.
And the best part is that NEAT doesn't require a powerful GPU.
If you do decide to do an evolutionary-based policy, I highly recommend messing around with NEAT (https://github.com/CodeReclaimers/neat-python).. I have successfully used it to play a number of SNES/Genesis games.
I've been using the python-neat library in open-ai's retro with some success. And while it works quickly, it normally finds local maximas. It seems to struggle with long sequences. And defining the fitness function/parameters is an artform.
Here's a video of Donkey Kong Country played by python-neat in open-ai's retro. It took 8 generations of 20 genones to beat level one. I'll post the code if anyone's interested.
They're rewarded by exploring, not by values in game.
Maybe alone they're not enough, but in conjunction with other things, I bet we could beat Zelda. I had the bot exploring enough to find the first dungeon of LoZ.