this is an old school runescape RL agent going against a completely unfair opponent, a perfect robot that always makes the "correct decision" and reads your actions before they happen 50% of the time. it's easy to tell which is which. to people who know anything about this game it might be a bit surprising but the policy mops the floor with the opponent. especially if you watch how it plays. it equips supposedly shit gear for what it's trying to do all the time, and barely does anything a top human player would want to do. it reaches easily 90% win rates within minutes of training on a single 4090. ofc the real test would be how it generalizes against the absolute best human players, but even though i have a human play mode here it's impossible for me to test since i get absolutely dumpstered by the policy immediately. i used to be really good at this, too. "just make it go against real players in the live game" is one consideration but we're not really eager to go straight up cheat at the game, would be very cool to get something private set up though!