Update.
[picoclvr.git] / problems.py
index 28e4f7b..ac16df4 100755 (executable)
@@ -24,6 +24,8 @@ class Problem:
 
 
 ####################
+
+
 class ProblemDegradation(Problem):
     def __init__(self, nb_state_tokens=5, nb_time_steps=12, value_max=25, hard=False):
         assert value_max // nb_state_tokens >= 2
@@ -287,18 +289,59 @@ class ProblemAddition(Problem):
 
 
 class ProblemMixing(Problem):
-    def __init__(self, height=3, width=3, nb_time_steps=12, hard=False):
+    def __init__(
+        self, height=4, width=4, nb_time_steps=9, hard=False, random_start=True
+    ):
         self.height = height
         self.width = width
         self.nb_time_steps = nb_time_steps
         self.hard = hard
+        self.random_start = random_start
+
+    def start_random(self, nb):
+        y = torch.arange(self.height * self.width).reshape(1, -1).expand(nb, -1)
+
+        if self.random_start:
+            i = (
+                torch.arange(self.height)
+                .reshape(1, -1, 1)
+                .expand(nb, self.height, self.width)
+            )
+            j = (
+                torch.arange(self.width)
+                .reshape(1, 1, -1)
+                .expand(nb, self.height, self.width)
+            )
+
+            ri = torch.randint(self.height, (nb,)).reshape(nb, 1, 1)
+            rj = torch.randint(self.width, (nb,)).reshape(nb, 1, 1)
+
+            m = 1 - torch.logical_or(i == ri, j == rj).long().flatten(1)
+
+            y = y * m + self.height * self.width * (1 - m)
+
+        y = y.reshape(nb, self.height, self.width)
 
-    def start(self, nb):
-        return (
-            torch.arange(self.height * self.width)
-            .reshape(1, 1, self.height, self.width)
-            .expand(nb, -1, -1, -1)
+        return y
+
+    def start_error(self, x):
+        i = torch.arange(self.height, device=x.device).reshape(1, -1, 1).expand_as(x)
+        j = torch.arange(self.width, device=x.device).reshape(1, 1, -1).expand_as(x)
+
+        ri = (
+            (x == self.height * self.width).long().sum(dim=-1).argmax(-1).view(-1, 1, 1)
         )
+        rj = (
+            (x == self.height * self.width).long().sum(dim=-2).argmax(-1).view(-1, 1, 1)
+        )
+
+        m = 1 - torch.logical_or(i == ri, j == rj).long().flatten(1)
+
+        x = x.flatten(1)
+        u = torch.arange(self.height * self.width, device=x.device).reshape(1, -1)
+
+        d = (x - (m * u + (1 - m) * self.height * self.width)).abs().sum(-1)
+        return d
 
     def moves(self, x):
         y = (
@@ -323,8 +366,7 @@ class ProblemMixing(Problem):
         return y
 
     def generate_sequences(self, nb):
-        y = self.start(nb)
-        x = y[torch.arange(nb), torch.randint(y.size(1), (nb,))]
+        x = self.start_random(nb)
 
         seq = [x.flatten(1)]
 
@@ -349,8 +391,7 @@ class ProblemMixing(Problem):
 
         x = a[0]
 
-        y = self.start(result.size(0)).to(x.device)
-        d = (x[:, None] - y).abs().sum((-1, -2)).min(dim=-1).values
+        d = self.start_error(x)
 
         for t in range(self.nb_time_steps - 1):
             x0, x = a[t], a[t + 1]
@@ -365,7 +406,15 @@ class ProblemMixing(Problem):
         return " | ".join(
             [
                 " ".join(
-                    ["-".join([f"{x:02d}" for x in s]) for s in r.split(self.width)]
+                    [
+                        "-".join(
+                            [
+                                f"{x:02d}" if x < self.height * self.width else "**"
+                                for x in s
+                            ]
+                        )
+                        for s in r.split(self.width)
+                    ]
                 )
                 for r in seq.split(self.height * self.width)
             ]
@@ -375,7 +424,7 @@ class ProblemMixing(Problem):
 ####################
 
 if __name__ == "__main__":
-    p = ProblemMixing(hard=True)
+    p = ProblemMixing()
     s, m = p.generate_sequences(10000)
     for x in s[:5]:
         print(p.seq2str(x))