- r = rewards
- print(f"{r.size()=} {lookahead_delta=}")
- u = F.pad(r, (0, lookahead_delta - 1)).as_strided(
- (r.size(0), r.size(1), lookahead_delta),
- (r.size(1) + lookahead_delta - 1, 1, 1),
- )
- a = u.min(dim=-1).values
- b = u.max(dim=-1).values
+ # r = rewards
+ # u = F.pad(r, (0, lookahead_delta - 1)).as_strided(
+ # (r.size(0), r.size(1), lookahead_delta),
+ # (r.size(1) + lookahead_delta - 1, 1, 1),
+ # )
+ # a = u[:, :, 1:].min(dim=-1).values
+ # b = u[:, :, 1:].max(dim=-1).values
+ # s = (a < 0).long() * a + (a >= 0).long() * b
+ # lookahead_rewards = (1 + s[:, :, None]) + first_lookahead_rewards_code
+
+ # a[n,t]=min_s>t r[n,s]
+ a = rewards.new_zeros(rewards.size())
+ b = rewards.new_zeros(rewards.size())
+ for t in range(a.size(1) - 1):
+ a[:, t] = rewards[:, t + 1 :].min(dim=-1).values
+ b[:, t] = rewards[:, t + 1 :].max(dim=-1).values