Home › In modern game AI, REINFORCEMENT LEARNING agents…
In modern game AI, REINFORCEMENT LEARNING agents (e.g. OpenAI Five, DeepMind AlphaStar) learn from which signal?
AHand-labelled per-action ground-truth dataset
BPure imitation of one expert recording always
CManually scripted FSMs by designers only
DCumulative reward from environment interactions
Answer & Solution
Correct answer: D. Cumulative reward from environment interactions
Sutton + Barto RL textbook + OpenAI / DeepMind blog posts: RL trains policy π to maximise expected discounted return Σ γ^t r_t from rollouts in the game environment.
Related questions
In Reynolds-style steering behaviours combined into a STEERING PIPELINE, the OUTPUT accelePer Dave Mark + Brian Schwab, a RESPONSE CURVE in utility AI maps an input variable to a sPer AlphaGo (Silver et al. 2016, Nature), MONTE CARLO TREE SEARCH (MCTS) explores a game tPer Daniel Harabor's follow-up work and the public literature, JPS+ (the offline variant) Per Daniel Harabor (2011 AAAI paper), JPS (Jump Point Search) accelerates A* on UNIFORM-COIn tactical AI, an INFLUENCE MAP is best described as which data structure?In behaviour trees beyond sequence + selector, a DECORATOR node typically does which?Per Dave Mark ('Behavioral Mathematics for Game AI'), UTILITY AI selects actions by which