RL without TD learning

Berkeley AI Research proposes an off-policy RL value-learning algorithm based on divide and conquer that no longer uses temporal difference (TD) learning, whose Bellman recursion error accumulation makes it hard to scale to long-horizon tasks.