Samovar

Version 0.2 (unreleased). Subject to change in backwards-incompatible ways.

Samovar is a DSL for modelling a world using propositions (facts), and possible events that can occur based on those facts, changing them.

Possible events are described with event rules. Each event rule looks like

[A] X [B]

and means "If A holds, then X is a possible action to take, and if you do take it, you must make B hold afterwards."

By "hold" we mean "matches the current set of facts."

As an example,

[actor(α),item(β),~holding(α,β)] α picks up the β. [holding(α,β)]

Which can be read

If α is an actor and β is an item and α is not holding β, then one possible action is to write out 'α picks up the β' and assert that α is now holding β.

We can add a complementary rule:

[actor(α),item(β),holding(α,β)] α puts down the β. [~holding(α,β)]

And we can package this all into a scenario:

scenario IgnatzWithBrick {

    [actor(α),item(β),~holding(α,β)]  α picks up the β.   [holding(α,β)]
    [actor(α),item(β),holding(α,β)]   α puts down the β.  [~holding(α,β)]

    actor(Ignatz).
    item(brick).

    goal [].
}

And an implementation of Samovar could take this scenario and use it to, among other things, generate textual descriptions of chains of events like

Ignatz picks up the brick. Ignatz puts down the brick.

Of course, this is a very simple example. A more complex example might have more actors, more items, and more rules (for example, that two actors cannot be holding the same item at the same time.)

Discussion

Samovar could be described as an "assertion-retraction engine", which itself could be thought of as a highly stylized form of Prolog programming plus some syntactic sugar.

Alternately, it could be thought of as assigning preconditions and postconditions, like you would find in Hoare logic, to actions in a world-model. Instead of proving that the action satisfies the conditions, though, we simply assume it does, and use the conditions to chain actions together in a sensible order.

But really, the internals are far simpler than an inference engine or a theorem prover: there are no logical rules in the database, only propositions, so they can be selected by simple pattern-matching rather than full unification.

TODO

• (+) Consider what it would take to add a predicate that evaluates to whether a given action has been taken previously or not.
• (+) Consider macros.
• Output scenarios to JSON.
• Python 3 support.
• Allow comments in source file. Probably # or //.
• Say that Greek letter variables are exactly equivalent to ?rho, etc.
• ! for not.
• Consider a simple equality rule.
• Consider allowing ∨.
• Take AST code from SixtyPical or ALPACA (or try to do better, perhaps with named tuples, but this is probably madness and not at all worth doing)

Probably we don't do

• (+) Allow sentence trees to be given for actions.
• (+) Allow scenarios to specify a minimum number of events to generate.