Allow for an increase in loss to find a simpler model
Source:R/trade_loss_for_simplicity.R
trade_loss_for_simplicity.RdAllow for an increase in loss to find a simpler model
Arguments
- alphas_grid
The data frame of model parameters being evaluated
- losses
A vector of length equal to rows of
alphas_grid, representing the loss associated with each set of model parameters- increase
The allowed increase in loss in percentage points compared to the best observed loss
Examples
# Returns `1` when all options are equal
alphas_grid <- data.frame(
alpha = rep(0, 10),
alpha_seasonal = rep(0, 10),
alpha_seasonal_decay = rep(0, 10)
)
losses <- rep(1, 10)
trade_loss_for_simplicity(
alphas_grid = alphas_grid,
losses = losses,
increase = 1
)
#> [1] 1
# Considers only options with 0 loss if best loss is 0
# (here, the originally best index with 0 loss has a less simple model)
alphas_grid <- data.frame(
alpha = c(0.5, rep(0, 2)),
alpha_seasonal = rep(0, 3),
alpha_seasonal_decay = rep(0, 3)
)
losses <- c(0, 1, 0)
trade_loss_for_simplicity(
alphas_grid = alphas_grid,
losses = losses,
increase = 1
)
#> [1] 3
# When multiple options exist in the allowed range of performance reduction,
# the best loss of those with the highest simplicity dominates
alphas_grid <- data.frame(
alpha = c(0.5, 0.5, 0.5, 0.5, 0),
alpha_seasonal = c(0.5, 0.5, 0, 0, 0),
alpha_seasonal_decay = c(0.5, 0, 0, 0, 0)
)
losses <- c(100, 101, 103, 102, 110)
trade_loss_for_simplicity(
alphas_grid = alphas_grid,
losses = losses,
increase = 5
)
#> [1] 4