Merge pull request #35 from Microsoft/parallel-quests-navigation

chaining: Re-generate actions when branching

Author: MarcCote

textworld/generator/chaining.py

@@ -76,6 +76,8 @@ class ChainingOptions:
         subquests:
             Whether to also return incomplete quests, which could be extended
             without reaching the depth or breadth limits.
+        independent_chains:
+            Whether to allow totally independent parallel chains.
         create_variables:
             Whether new variables may be created during chaining.
         fixed_mapping:
@@ -99,6 +101,7 @@ def __init__(self):
         self.min_breadth = 1
         self.max_breadth = 1
         self.subquests = False
+        self.independent_chains = False
         self.create_variables = False
         self.fixed_mapping = data.get_types().constants_mapping
         self.rng = None
@@ -189,30 +192,17 @@ class _Node:
     A node in a chain being generated.
 
     Each node is aware of its position (depth, breadth) in the dependency tree
-    induced by the chain.  For generating parallel quests, the backtracks field
-    holds actions that can be use to go up the dependency tree and start a new
-    chain.
-
-    For example, taking the action node.backtracks[i][j] will produce a new node
-    at depth (i + 1) and breadth (self.breadth + 1).  To avoid duplication, in
-    trees like this:
-
-         root
-         / | \
-        A  B  C
-        |  |  |
-        .......
-
-    A.backtracks[0] will be [B, C], B.backtracks[0] will be [C], and
-    C.backtracks[0] will be [].
+    induced by the chain.  To avoid duplication when generating parallel chains,
+    each node stores the actions that have already been used at that depth.
     """
 
-    def __init__(self, parent, dep_parent, state, action, backtracks, depth, breadth):
+    def __init__(self, parent, dep_parent, state, action, rules, used, depth, breadth):
         self.parent = parent
         self.dep_parent = dep_parent
         self.state = state
         self.action = action
-        self.backtracks = backtracks
+        self.rules = rules
+        self.used = used
         self.depth = depth
         self.breadth = breadth
 
@@ -235,7 +225,7 @@ def __init__(self, state, options):
 
     def root(self) -> _Node:
         """Create the root node for chaining."""
-        return _Node(None, None, self.state, None, [], 0, 1)
+        return _Node(None, None, self.state, None, [], set(), 0, 1)
 
     def chain(self, node: _Node) -> Iterable[_Node]:
         """
@@ -251,30 +241,21 @@ def chain(self, node: _Node) -> Iterable[_Node]:
         if self.rng:
             self.rng.shuffle(assignments)
 
-        partials = []
-        actions = []
-        states = []
+        used = set()
         for partial in assignments:
             action = self.try_instantiate(node.state, partial)
             if not action:
                 continue
 
-            if not self.check_action(node, action):
+            if not self.check_action(node, node.state, action):
                 continue
 
             state = self.apply(node, action)
             if not state:
                 continue
 
-            partials.append(partial)
-            actions.append(action)
-            states.append(state)
-
-        for i, action in enumerate(actions):
-            # Only allow backtracking into later actions, to avoid duplication
-            remaining = partials[i+1:]
-            backtracks = node.backtracks + [remaining]
-            yield _Node(node, node, states[i], action, backtracks, node.depth + 1, node.breadth)
+            used = used | {action}
+            yield _Node(node, node, state, action, rules, used, node.depth + 1, node.breadth)
 
     def backtrack(self, node: _Node) -> Iterable[_Node]:
         """
@@ -284,21 +265,39 @@ def backtrack(self, node: _Node) -> Iterable[_Node]:
         if node.breadth >= self.max_breadth:
             return
 
-        for i, partials in enumerate(node.backtracks):
-            backtracks = node.backtracks[:i]
-
-            for j, partial in enumerate(partials):
+        parent = node
+        parents = []
+        while parent.dep_parent:
+            if parent.depth == 1 and not self.options.independent_chains:
+                break
+            parents.append(parent)
+            parent = parent.dep_parent
+        parents = parents[::-1]
+
+        for sibling in parents:
+            parent = sibling.dep_parent
+            rules = self.options.get_rules(parent.depth)
+            assignments = self.all_assignments(node, rules)
+            if self.rng:
+                self.rng.shuffle(assignments)
+
+            for partial in assignments:
                 action = self.try_instantiate(node.state, partial)
                 if not action:
                     continue
 
+                if action in sibling.used:
+                    continue
+
+                if not self.check_action(parent, node.state, action):
+                    continue
+
                 state = self.apply(node, action)
                 if not state:
                     continue
 
-                remaining = partials[j+1:]
-                new_backtracks = backtracks + [remaining]
-                yield _Node(node, partial.node, state, action, new_backtracks, i + 1, node.breadth + 1)
+                used = sibling.used | {action}
+                yield _Node(node, parent, state, action, rules, used, sibling.depth, node.breadth + 1)
 
     def all_assignments(self, node: _Node, rules: Iterable[Rule]) -> Iterable[_PartialAction]:
         """
@@ -359,7 +358,7 @@ def create_variable(self, state, ph, type_counts):
         type_counts[ph.type] += 1
         return var
 
-    def check_action(self, node: _Node, action: Action) -> bool:
+    def check_action(self, node: _Node, state: State, action: Action) -> bool:
         # Find the last action before a navigation action
         # TODO: Fold this behaviour into ChainingOptions.check_action()
         nav_parent = node
@@ -387,7 +386,7 @@ def check_action(self, node: _Node, action: Action) -> bool:
             if len(recent.added & relevant) == 0 or len(pre_navigation.added & relevant) == 0:
                 return False
 
-        return self.options.check_action(node.state, action)
+        return self.options.check_action(state, action)
 
     def _is_navigation(self, action):
         return action.name.startswith("go/")
@@ -405,8 +404,8 @@ def apply(self, node: _Node, action: Action) -> Optional[State]:
 
         new_state.apply(action)
 
-        # Some debug checks
-        assert self.check_state(new_state)
+        if not self.check_state(new_state):
+            return None
 
         # Detect cycles
         state = new_state.copy()

textworld/generator/tests/test_chaining.py

@@ -237,4 +237,73 @@ def test_parallel_quests():
     options.min_breadth = 1
     options.create_variables = True
     chains = list(get_chains(State(), options))
-    assert len(chains) == 6
+    assert len(chains) == 5
+
+
+def test_parallel_quests_navigation():
+    logic = GameLogic.parse("""
+        type P {
+        }
+
+        type I {
+        }
+
+        type r {
+            rules {
+                move :: at(P, r) & $free(r, r') -> at(P, r');
+            }
+
+            constraints {
+                atat :: at(P, r) & at(P, r') -> fail();
+            }
+        }
+
+        type o {
+            rules {
+                take :: $at(P, r) & at(o, r) -> in(o, I);
+            }
+
+            constraints {
+                inat :: in(o, I) & at(o, r) -> fail();
+            }
+        }
+
+        type flour : o {
+        }
+
+        type eggs : o {
+        }
+
+        type cake {
+            rules {
+                bake :: in(flour, I) & in(eggs, I) -> in(cake, I) & in(flour, cake) & in(eggs, cake);
+            }
+
+            constraints {
+                inincake :: in(o, I) & in(o, cake) -> fail();
+                atincake :: at(o, r) & in(o, cake) -> fail();
+            }
+        }
+    """)
+
+    state = State([
+        Proposition.parse("at(P, r3: r)"),
+        Proposition.parse("free(r2: r, r3: r)"),
+        Proposition.parse("free(r1: r, r2: r)"),
+    ])
+
+    bake = [logic.rules["bake"]]
+    non_bake = [r for r in logic.rules.values() if r.name != "bake"]
+
+    options = ChainingOptions()
+    options.backward = True
+    options.create_variables = True
+    options.min_depth = 3
+    options.max_depth = 3
+    options.min_breadth = 2
+    options.max_breadth = 2
+    options.logic = logic
+    options.rules_per_depth = [bake, non_bake, non_bake]
+    options.restricted_types = {"P", "r"}
+    chains = list(get_chains(state, options))
+    assert len(chains) == 2