Update.

diff --git a/escape.py b/escape.py
index aa87cdd..a8b9536 100755 (executable)
--- a/escape.py
+++ b/escape.py
@@ -12,13 +12,15 @@ from torch.nn import functional as F
 ######################################################################
 
 nb_state_codes = 4
-nb_rewards_codes = 3
 nb_actions_codes = 5
+nb_rewards_codes = 3
+nb_lookahead_rewards_codes = 3
 
 first_state_code = 0
-first_rewards_code = first_state_code + nb_state_codes
-first_actions_code = first_rewards_code + nb_rewards_codes
-nb_codes = first_actions_code + nb_actions_codes
+first_actions_code = first_state_code + nb_state_codes
+first_rewards_code = first_actions_code + nb_actions_codes
+first_lookahead_rewards_code = first_rewards_code + nb_rewards_codes
+nb_codes = first_lookahead_rewards_code + nb_lookahead_rewards_codes
 
 ######################################################################
 
@@ -103,10 +105,24 @@ def generate_episodes(nb, height=6, width=6, T=10):
 ######################################################################
 
 
-def episodes2seq(states, actions, rewards):
+def episodes2seq(states, actions, rewards, lookahead_delta=None):
     states = states.flatten(2) + first_state_code
     actions = actions[:, :, None] + first_actions_code
-    rewards = (rewards[:, :, None] + 1) + first_rewards_code
+
+    if lookahead_delta is not None:
+        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
+        s = (a < 0).long() * a + (a >= 0).long() * b
+        lookahead_rewards = (1 + s[:, :, None]) + first_lookahead_rewards_code
+
+    r = rewards[:, :, None]
+    rewards = (r + 1) + first_rewards_code
 
     assert (
         states.min() >= first_state_code
@@ -121,29 +137,48 @@ def episodes2seq(states, actions, rewards):
         and rewards.max() < first_rewards_code + nb_rewards_codes
     )
 
-    return torch.cat([states, actions, rewards], dim=2).flatten(1)
+    if lookahead_delta is None:
+        return torch.cat([states, actions, rewards], dim=2).flatten(1)
+    else:
+        assert (
+            lookahead_rewards.min() >= first_lookahead_rewards_code
+            and lookahead_rewards.max()
+            < first_lookahead_rewards_code + nb_lookahead_rewards_codes
+        )
+        return torch.cat([states, actions, rewards, lookahead_rewards], dim=2).flatten(
+            1
+        )
 
 
-def seq2episodes(seq, height, width):
-    seq = seq.reshape(seq.size(0), -1, height * width + 2)
+def seq2episodes(seq, height, width, lookahead=False):
+    seq = seq.reshape(seq.size(0), -1, height * width + (3 if lookahead else 2))
     states = seq[:, :, : height * width] - first_state_code
     states = states.reshape(states.size(0), states.size(1), height, width)
     actions = seq[:, :, height * width] - first_actions_code
     rewards = seq[:, :, height * width + 1] - first_rewards_code - 1
-    return states, actions, rewards
+
+    if lookahead:
+        lookahead_rewards = (
+            seq[:, :, height * width + 2] - first_lookahead_rewards_code - 1
+        )
+        return states, actions, rewards, lookahead_rewards
+    else:
+        return states, actions, rewards
 
 
 ######################################################################
 
 
-def episodes2str(states, actions, rewards, unicode=False, ansi_colors=False):
+def episodes2str(
+    states, actions, rewards, lookahead_rewards=None, unicode=False, ansi_colors=False
+):
     if unicode:
         symbols = " █@$"
         # vert, hori, cross, thin_hori = "║", "═", "╬", "─"
-        vert, hori, cross, thin_hori = "┃", "━", "╋", "─"
+        vert, hori, cross, thin_vert, thin_hori = "┃", "━", "╋", "│", "─"
     else:
         symbols = " #@$"
-        vert, hori, cross, thin_hori = "|", "-", "+", "-"
+        vert, hori, cross, thin_vert, thin_hori = "|", "-", "+", "|", "-"
 
     hline = (cross + hori * states.size(-1)) * states.size(1) + cross + "\n"
 
@@ -167,11 +202,14 @@ def episodes2str(states, actions, rewards, unicode=False, ansi_colors=False):
 
         result += (vert + thin_hori * states.size(-1)) * states.size(1) + vert + "\n"
 
-        def status_bar(a, r):
-            a = a.item()
-            a = "ISNEW"[a] if a >= 0 and a < 5 else "?"
-            r = "?" if r < -1 or r > 2 else ("" if r == 0 else f"{r.item()}")
-            return a + " " * (states.size(-1) - len(a) - len(r)) + r
+        def status_bar(a, r, lr=None):
+            a, r = a.item(), r.item()
+            sb = "ISNEW"[a] if a >= 0 and a < 5 else "?"
+            sb = sb + thin_vert + ("- +"[r + 1] if r in {-1, 0, 1} else "?")
+            if lr is not None:
+                lr = lr.item()
+                sb = sb + thin_vert + +("- +"[lr + 1] if lr in {-1, 0, 1} else "?")
+            return sb + " " * (states.size(-1) - len(sb))
 
         result += (
             vert
@@ -194,6 +232,6 @@ def episodes2str(states, actions, rewards, unicode=False, ansi_colors=False):
 if __name__ == "__main__":
     nb, height, width, T = 8, 4, 6, 20
     states, actions, rewards = generate_episodes(nb, height, width, T)
-    seq = episodes2seq(states, actions, rewards)
-    s, a, r = seq2episodes(seq, height, width)
-    print(episodes2str(s, a, r, unicode=True, ansi_colors=True))
+    seq = episodes2seq(states, actions, rewards, lookahead_delta=2)
+    s, a, r, lr = seq2episodes(seq, height, width, lookahead=True)
+    print(episodes2str(s, a, r, lookahead_rewards=lr, unicode=True, ansi_colors=True))