Skip to content

Latest commit

 

History

History
221 lines (194 loc) · 7.47 KB

scone.md

File metadata and controls

221 lines (194 loc) · 7.47 KB

Executor language for the SCONE domains

This file describes the program syntax for the executor implemented in strongsup/rlong/executor.py. This is the executor used in "From Language to Programs: Bridging Reinforcement Learning and Maximum Marginal Likelihood".

Note #1: the history referencing predicates mentioned in the paper are named slightly differently in this implementation:

  • prevAction = H0
  • prevArg1 = H1
  • prevArg2 = H2

Note #2: you may see several files/classes in the project which mention rlong. These are associated with the SCONE dataset, which was created by Reginald Long.

Overview

  • World state
    • A world state is a list of objects.
    • Each object has several properties (colors, position, etc.).
  • Logical form
    • Each utterance maps to a single command.
    • Each command contains a single action and several arguments.
    • The arguments are built up before an action is performed.
  • Execution
    • The executor processes the logical form from left to right.
    • The executed denotation contains:
      • world_state (initialized to a specified initial state)
      • command_history (initially an empty list)
      • utterance_idx (initially 0)
      • denotation_stack (initially an empty stack)
    • Most logical form predicates will modify only the denotation_stack.
    • The world_state, command_history, and utterance_idx are modified when an action is executed (as explained in the next section)
    • A logical form is terminated when utterance_idx exceeds the bound of the list of utterances (i.e., utterance_idx == len(context.utterances)). The finalized denotation is the final world_state.

Predicates and execution

  • Primitives

    • The primitives include:

      • Positive numbers: 1, 2, 3, ...
      • Negative numbers: -1, -2, -3, ...
      • Fractions (prefix: X): X1/1, X1/2, X3/4, ...
      • Colors (single char): r, g, o, y, p, b, e

    • Execution: Push the primitive onto the stack

    • Examples:

      lf              | stack
3               | 3
3 g             | 3 g
3 g X1/3        | 3 g X1/3

  • Properties (prefix: P): PColor, PShirt, ...

    • Execution:

      • Pop v from the stack
      • Perform some computation (e.g., look up objects with that property v) and push the result onto the stack

    • Examples:

      world_state: 1:g 2: 3:yyy 4:b 5:gg
lf              | stack
g               | g
g PColor        | [1:g, 5:gg]

  • Double-Properties (prefix: D): DShirtHat, ...

    • Execution:

      • Pop 2 things from the stack
      • Perform some computation (e.g., look up objects with those 2 property values) and push the result onto the stack

    • Examples: (Scene domain)

      world_state: 1:__ 2:__ 3:__ 4:__ 5:gr 6:__ 7:go 8:__ 9:__ 10:yo
lf              | stack
g o             | g o
g o DShirtHat   | [7:go]

  • all-objects

    • Execution: Push the list of all objects onto the stack

    • Examples:

      world_state: 1:g 2: 3:yyy 4:b 5:gg
lf              | stack
all-objects     | [1:g, 2:, 3:yyy, 4:b, 5:gg]

  • index

    • Execution:

      • Pop a list and a number i (positive or negative) from the stack
      • Push the list member with index i onto the stack

    • Examples:

      world_state: 1:g 2: 3:yyy 4:b 5:gg
lf                  | stack
all-objects 4       | [1:g, 2:, 3:yyy, 4:b, 5:gg] 4
all-objects 4 index | [4:b]
g PColor            | [1:g, 5:gg]
g PColor -1         | [1:g, 5:gg] -1
g PColor -1 index   | [5:gg]

  • Actions (prefix: A): APour, ADrain, ARemove, ...

    • Execution:

      • Pop the action arguments from the stack. The action determines the number of arguments. Optional requirement: The stack should become empty
      • Apply the action on the world state
      • Save the command (action and arguments) to the command history
      • Increment the utterance count by 1

    • Examples:

      world_state: 1:g 2: 3:yyy 4:b 5:gg
lf                             | stack
all-objects 4 index 1          | [4:b] 1
all-objects 4 index 1 ADrain   | 
    new world_state:     1:g 2: 3:yyy 4: 5:gg
    new command_history: [(Drain, 4:b, 1)]
    new utterance_idx:   1

  • History (prefix: H): H0, H1, H2, ...

    • Execution:

      • Pop a command index i from the stack
      • Look at the ith command history entry. If the predicate is H0, pull out the action and immediately execute it. Otherwise, for predicate Hj, pull out the jth argument and push it onto the stack.
        • If it is an object, retrieve the latest attributes for that object
        • If the object has left the scene, then use the attributes that were present when the command was executed.

    • Examples:

      (After executing "all-objects 4 index 1 ADrain")
world_state: 1:g 2: 3:yyy 4: 5:gg
command_history: [(Drain, 4:b, 1)]
lf                   | stack
1 H1                 | 4:b         # 1st argument of the first command
-1 H2                | 1           # 2nd argument of the last command
y PColor -1 H2       | [3:yyy] 1
y PColor -1 H2 -1 H0 |             # Invoke the action of the last command
    new world_state:     1:g 2: 3:yy 4: 5:gg
    new command_history: [(Drain, 4:b, 1), (Drain, 3:yyy 1)]
    new utterance_idx:   2

Subdomains

Alchemy

  • World state:
    • 7 beakers
    • Each beaker has up to 4 units of chemical
    • Each unit of chemical has a color (r, g, y, p, o, b)
  • Properties:
    • COLOR PColor: Selects all beakers with a homogeneous content of the specified color.
  • Actions:
    • OBJECT POSITIVE_NUMBER|FRACTION ADrain: Remove the specified amount of chemical from the beaker OBJECT
    • OBJECT1 OBJECT2 APour: Pour all chemical from OBJECT1 to OBJECT2
    • OBJECT AMix: Mix the chemical in OBJECT, making it turn brown (b).

Scene

  • World state:
    • A linear stage with 10 slots
    • Each slot may have a single person standing on it
    • Each person has a shirt and optionally a hat. Both the shirt and the hat has a color.
    • (Internal) Each person has a hidden ID. When H1, H2, ... is executed, the person is retrieved based on the ID.
  • Properties:
    • COLOR PShirt: Selects all people with the specified shirt color
    • COLOR|e PHat: Selects all people with the specified hat color
    • OBJECT PLeft: Returns the position to the left of the object
    • OBJECT PRight: Returns the position to the right of the object
    • COLOR COLOR|e DShirtHat: Selects all people with the specified shirt and hat color
  • Actions:
    • OBJECT ALeave: Remove the specified person OBJECT
    • OBJECT1 OBJECT2 ASwapHats: Swap the hats of OBJECT1 and OBJECT2
    • OBJECT NUMBER AMove: Move OBJECT to the specified absolute location
    • NUMBER COLOR COLOR|e ACreate: Create a new person at the specified position, shirt color, and hat color. Special: The saved command history is ('Create', person)

Tangrams

  • World state:
    • A row of tangrams with shape encoded as 0, 1, 2, ... The shape, which is also used as a unique ID, cannot be referred to directly.
  • Properties:
    • (None)
  • Actions:
    • NUMBER OBJECT AAdd: Insert a tangram to the specified location
    • OBJECT1 OBJECT2 ASwap: Swap the positions of two tangrams (specified as object or index)
    • OBJECT ARemove: Remove a tangram (specified as object or index)