Adaptive sequencing allocation via Thompson sampling

Reallocates sequencing resources across regions in real time, directing effort toward strata where uncertainty reduction has the highest decision value. Unlike static Neyman allocation (which optimises for a single time point), this function adapts to evolving variant dynamics across multiple surveillance rounds.

Usage

adaptive_design(
  data,
  capacity,
  n_rounds = NULL,
  strategy = c("thompson", "ucb"),
  target_lineage = NULL,
  exploration = 2,
  seed = NULL
)

Arguments

data

An lfq_data object with a location column.

capacity

Total sequencing capacity per round (integer).

n_rounds

Number of allocation rounds to simulate. Default NULL uses the number of time points in the data.

strategy

Allocation strategy: "thompson" (default) for Thompson sampling, or "ucb" for Upper Confidence Bound.

target_lineage

Character; lineage to optimise detection for. Default NULL optimises for overall frequency estimation.

exploration

Exploration parameter for UCB. Default 2.0. Larger values explore more; smaller values exploit current best regions.

seed

Random seed. Default NULL.

Value

An adaptive_allocation S3 class with components:

allocations

Tibble with round, region, n_allocated, uncertainty, frequency.

summary

Tibble with per-region totals and mean allocation.

strategy

Character; strategy used.

capacity

Integer; per-round capacity.

Details

Thompson sampling draws from the posterior distribution of frequency estimates and allocates proportional to the posterior variance. Regions with high uncertainty receive more sequences, but the stochastic draws naturally balance exploration (sampling uncertain regions) and exploitation (sampling where variants are most prevalent).

UCB allocates proportional to the upper confidence bound of the estimation error: \(\text{score}_r = \hat{\sigma}_r + c \sqrt{2 \log(t) / n_r}\) where \(\hat{\sigma}_r\) is the current estimation uncertainty in region \(r\), \(n_r\) is the cumulative allocation, \(t\) is the round, and \(c\) is the exploration parameter.

See also

surveillance_value for EVOI analysis.

Examples

# \donttest{
sim <- simulate_dynamics(n_lineages = 3,
  advantages = c("A" = 1.3, "B" = 0.9),
  n_timepoints = 12, seed = 1)
ad <- adaptive_design(sim, capacity = 200, n_rounds = 8)
ad
#> 
#> ── Adaptive sequencing allocation 
#> Strategy: thompson | Capacity: 200/round
#> 8 rounds, 5 regions
#> 
#> # A tibble: 5 × 4
#>   region   total_allocated mean_allocated mean_uncertainty
#>   <chr>              <dbl>          <dbl>            <dbl>
#> 1 region_5             344           43            0.00105
#> 2 region_4             324           40.5          0.00105
#> 3 region_2             317           39.6          0.00105
#> 4 region_1             308           38.5          0.00105
#> 5 region_3             307           38.4          0.00105
# }