Dsxir.Optimizer.SIMBA.Trial (dsxir v0.5.0)

One SIMBA mini-batch step over a Dsxir.Optimizer.SIMBA.Sampler.

run/1 draws the next mini-batch (reshuffle-and-wrap on exhaustion), samples diverse trajectories from the current pool, buckets them per example, builds new candidate programs by dropping demos and applying a strategy, evaluates the candidates on the same mini-batch, records the per-step winner, and registers every candidate back into the pool.

All randomness threads sampler.rng_state explicitly: the batch reshuffle, softmax program selection, the Poisson demo-drop count, the dropped indices, and the strategy choice all take and return the :rand state. Nothing touches global :rand, so a resumed checkpoint replays deterministically.

Summary

Types

args()

Functions

drop_demos(program, config, rng)

Drops a Poisson-sampled number of demo positions from every predictor of program, threading rng.

run(args)

Runs one mini-batch step and returns {:ok, updated_sampler, trial_result}.

Types

args()

@type args() :: %{
  :sampler => Dsxir.Optimizer.SIMBA.Sampler.t(),
  :trial_idx => non_neg_integer(),
  :metric => Dsxir.Metric.t(),
  optional(:opts) => keyword()
}

Functions

drop_demos(program, config, rng)

@spec drop_demos(Dsxir.Program.t(), map(), term()) :: {Dsxir.Program.t(), term()}

Drops a Poisson-sampled number of demo positions from every predictor of program, threading rng.

num_demos is the maximum demo count across predictors. With max_demos_tmp = max(max_demos, 3 when max_demos == 0), the drop count is min(max(poisson(num_demos / max_demos_tmp), floor), num_demos) where floor is 1 when num_demos >= max_demos_tmp and 0 otherwise. The dropped positions are sampled with replacement and removed from each predictor's State.demos.

run(args)

@spec run(args()) :: {:ok, Dsxir.Optimizer.SIMBA.Sampler.t(), map()}

Runs one mini-batch step and returns {:ok, updated_sampler, trial_result}.

The updated sampler carries the threaded rng_state, the advanced instance_idx (and reshuffled data_indices on wrap), the grown program pool / score lists / winning programs, the updated best_so_far, an incremented attempts, and the per-step entry under trial_logs[trial_idx].