Open AchillesJJ opened 6 years ago
However, it seems that such loss function only contains instantaneous reward, not average cumulated reward.
If there is no commission fee, when the action won't affect the state transition, optimizing the immediate rewards is equivalent to optimizing the long-term value. And this point, together with the differentiable reward function, gives superior sample efficiency compared with common purpose RL. To deal with the commission fee, we treat it as a regularization term.
As show in the
nnagent.py
, the author use average return of a batch as the loss function. However, it seems that such loss function only contains instantaneous reward, not average cumulated reward. To be specific, supposing we have a batch of experience as followsmini_batch = $(s_t, a_t, rt, ..., s(t+T), a(t+T), r(t+T))$