Doodle

i taught a bot to play doodle jump

google interview prep, but make it playable: a doodle jump clone where a reinforcement learning bot learns to play and you watch it go from flailing to competent. the part i like: everything is explicit. reward = 0.55 x replay + 0.35 x session length + 0.10 x golden-snitch share. updates capped at 5% per run. bots retest every 50 runs and roll back versions that got worse. separate phone-tilt and keyboard contexts. policy, exploration, baselines, advantage, policy gradients, all visible in a game instead of equations. try it: savarsareen.com/doodle