Smart Handlers: Handling the selection monad
summary
Introduces selection monad ideas into algebraic effect handlers to add choice continuations that compute future losses; prototypes in a Haskell-like Huscoll library and a planned JAX embedding. Claims handlers can implement optimizers (greedy pick chooses "hello world" with loss −93 vs ICFP25 −72), SGD and hyperparameter tuning as handlers, and two-player minimax strategies. Formal work includes an operational calculus, proofs of progress/preservation and denotational soundness/adequacy; implementation is a working prototype.
AI-written summary · about
chapters (14)
- 0:00Introduction
- 6:20Algebraic effect handlers
- 10:17Password selection with effects
- 13:55Selection monad
- 23:29Smart handlers overview
- 27:46Calculus syntax and types
- 32:46Smart handlers semantics
- 43:25Password selection evaluation
- 47:47Linear regression with handlers
- 53:13Hyperparameter tuning
- 56:05Minimax strategies in games
- 1:00:01Termination and semantics
- 1:01:04Future work and Jax
- 1:04:00Q&A
related talks
21mHandling the Selection Monad
15mEffect Handlers, Evidently
26mEffect Handlers for Programmable Inference
25mHandling Higher-Order Effectful Operations with Judgemental Monadic Laws
29mStaging Effect Handlers for Modular Search
17mHefty Algebras: Modular Elaboration of Higher-Order Algebraic Effects
22mType-Safe Code Generation With Algebraic Effects and Handlers
28mFirst-class Names for Effect Handlers