The main difficulty is in how many hyperparameters are involved in training an RL agent, high sensitivity of RL algorithms to those hyperparameters, and not having a good understanding of how to select them based on the properties of your task. This problem is exacerbated by the high sample complexity of RL. If something doesn’t work out, you don’t know if it’s because you chose the wrong set of hyperparameters or if you just haven’t trained for long enough.
I don’t know much about game design, but I do know that it’s a much more mature field than RL, so surely they have better tools than guessing and praying.
I don’t know if you can describe it as “can’t be arsed” when their proposed solution is so much harder to implement.
Opening an rdp connection to some dude in India so he can do it for you isn’t that hard.
Like some other AIs ended up being.
Is it? What’s so fucking hard about it?
The main difficulty is in how many hyperparameters are involved in training an RL agent, high sensitivity of RL algorithms to those hyperparameters, and not having a good understanding of how to select them based on the properties of your task. This problem is exacerbated by the high sample complexity of RL. If something doesn’t work out, you don’t know if it’s because you chose the wrong set of hyperparameters or if you just haven’t trained for long enough.
I don’t know much about game design, but I do know that it’s a much more mature field than RL, so surely they have better tools than guessing and praying.