Update.

[picoclvr.git] / escape.py

diff --git a/escape.py b/escape.py
index f2a1662..5f34cd1 100755 (executable)
--- a/escape.py
+++ b/escape.py
@@ -11,10 +11,10 @@ from torch.nn import functional as F
 
 ######################################################################
 
-nb_states_codes = 4
+nb_states_codes = 5
 nb_actions_codes = 5
 nb_rewards_codes = 3
-nb_lookahead_rewards_codes = 3
+nb_lookahead_rewards_codes = 4  # stands for -1, 0, +1, and UNKNOWN
 
 first_states_code = 0
 first_actions_code = first_states_code + nb_states_codes
@@ -50,6 +50,7 @@ def code2reward(r):
 
 
 def lookahead_reward2code(r):
+    # -1, 0, +1 or 2 for UNKNOWN
     return r + 1 + first_lookahead_rewards_code
 
 
@@ -60,14 +61,13 @@ def code2lookahead_reward(r):
 ######################################################################
 
 
-def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3):
+def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3, nb_coins=2):
     rnd = torch.rand(nb, height, width)
     rnd[:, 0, :] = 0
     rnd[:, -1, :] = 0
     rnd[:, :, 0] = 0
     rnd[:, :, -1] = 0
     wall = 0
-
     for k in range(nb_walls):
         wall = wall + (
             rnd.flatten(1).argmax(dim=1)[:, None]
@@ -76,6 +76,17 @@ def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3):
 
         rnd = rnd * (1 - wall.clamp(max=1))
 
+    rnd = torch.rand(nb, height, width)
+    coins = torch.zeros(nb, T, height, width, dtype=torch.int64)
+    rnd = rnd * (1 - wall.clamp(max=1))
+    for k in range(nb_coins):
+        coins[:, 0] = coins[:, 0] + (
+            rnd.flatten(1).argmax(dim=1)[:, None]
+            == torch.arange(rnd.flatten(1).size(1))[None, :]
+        ).long().reshape(rnd.size())
+
+        rnd = rnd * (1 - coins[:, 0].clamp(max=1))
+
     states = wall[:, None, :, :].expand(-1, T, -1, -1).clone()
 
     agent = torch.zeros(states.size(), dtype=torch.int64)
@@ -131,9 +142,12 @@ def generate_episodes(nb, height=6, width=6, T=10, nb_walls=3):
 
         # assert hit.min() == 0 and hit.max() <= 1
 
-        rewards[:, t + 1] = -hit + (1 - hit) * agent[:, t + 1, -1, -1]
+        got_coin = (agent[:, t + 1] * coins[:, t]).flatten(1).sum(dim=1)
+        coins[:, t + 1] = coins[:, t] * (1 - agent[:, t + 1])
+
+        rewards[:, t + 1] = -hit + (1 - hit) * got_coin
 
-    states += 2 * agent + 3 * monster
+    states = states + 2 * agent + 3 * monster + 4 * coins
 
     return states, agent_actions, rewards
 
@@ -173,7 +187,7 @@ def seq2episodes(seq, height, width):
 def seq2str(seq):
     def token2str(t):
         if t >= first_states_code and t < first_states_code + nb_states_codes:
-            return " #@$"[t - first_states_code]
+            return " #@T$"[t - first_states_code]
         elif t >= first_actions_code and t < first_actions_code + nb_actions_codes:
             return "ISNEW"[t - first_actions_code]
         elif t >= first_rewards_code and t < first_rewards_code + nb_rewards_codes:
@@ -182,7 +196,7 @@ def seq2str(seq):
             t >= first_lookahead_rewards_code
             and t < first_lookahead_rewards_code + nb_lookahead_rewards_codes
         ):
-            return "n.p"[t - first_lookahead_rewards_code]
+            return "n.pU"[t - first_lookahead_rewards_code]
         else:
             return "?"
 
@@ -196,11 +210,11 @@ def episodes2str(
     lookahead_rewards, states, actions, rewards, unicode=False, ansi_colors=False
 ):
     if unicode:
-        symbols = "·█@$"
+        symbols = "·█@T$"
         # vert, hori, cross, thin_hori = "║", "═", "╬", "─"
         vert, hori, cross, thin_vert, thin_hori = "┃", "━", "╋", "│", "─"
     else:
-        symbols = " #@$"
+        symbols = " #@T$"
         vert, hori, cross, thin_vert, thin_hori = "|", "-", "+", "|", "-"
 
     hline = (cross + hori * states.size(-1)) * states.size(1) + cross + "\n"
@@ -228,7 +242,7 @@ def episodes2str(
         def status_bar(a, r, lr=None):
             a, r = a.item(), r.item()
             sb_a = "ISNEW"[a] if a >= 0 and a < 5 else "?"
-            sb_r = "- +"[r + 1] if r in {-1, 0, 1} else "?"
+            sb_r = "- +U"[r + 1] if r in {-1, 0, 1, 2} else "?"
             if lr is None:
                 sb_lr = ""
             else:
@@ -266,7 +280,7 @@ def episodes2str(
 ######################################################################
 
 if __name__ == "__main__":
-    nb, height, width, T, nb_walls = 5, 5, 7, 20, 5
+    nb, height, width, T, nb_walls = 5, 5, 7, 10, 5
     states, actions, rewards = generate_episodes(nb, height, width, T, nb_walls)
     seq = episodes2seq(states, actions, rewards)
     lr, s, a, r = seq2episodes(seq, height, width)