decideR

decideR is the shared Bayesian decision layer of the ORCHESTRA analytics framework. It maps a posterior into an action under an explicit utility and a hard feasibility constraint, returning the action together with a grounding label and an abstention when the evidence will not support a decision. The grounding label is combined worst-case across inputs, so a decision resting on an unverified external fact is downgraded to abstention -- the Independent Oracle Principle -- rather than asserted with false confidence.

The same engine sizes a crop nitrogen rate and a trading position, because both are the same problem: act under uncertainty, and refuse to act when the evidence will not support it.

As the tail of an ORCHESTRA pipeline it decides directly from an upstream member's orchestra_manifest -- a PESTO inversion, a kernR causal test, a flexyBayes posterior -- via decide_from_manifest(), reading the predictive draws and the producer's grounding token off the contract without depending on the contract package. Its loss library includes a grade-band loss (grade_band_value()) that prices quality-threshold payoffs -- a grain price that steps up as protein clears a milling and a premium band -- as a first-class step value schedule.

Related packages

• decisionSupport chooses between two alternatives by expected welfare from Monte Carlo simulation, and uses value-of-information analysis to decide which uncertainties are worth reducing.
• BCEA takes cost and effectiveness draws for two or more interventions, for example MCMC output from a Bayesian model, and finds the most cost-effective one with probabilistic sensitivity analysis.
• voi calculates the expected value of perfect, partial and sample information from a decision model.

decideR also starts from posterior draws, but chooses from a range of actions (a nitrogen rate, a position size) under an explicit utility and a hard constraint, can price stepped payoffs such as grain protein bands, and abstains instead of deciding when an input is not grounded.

Installation

pak::pak("max578/decideR")

Example

The two reference decisions, from two domains, both route through the one engine.

library(decideR)

# A crop decision: choose the nitrogen rate of greatest expected profit, given
# posterior yield draws at each candidate rate.
rates <- seq(0, 200, by = 25)
ymax  <- rnorm(2000L, 5.5, 0.3)
yield_draws <- vapply(rates,
  function(r) 3 + (ymax - 3) * (1 - exp(-r / 80)) + rnorm(2000L, 0, 0.1),
  numeric(2000L))
decide_input_rate(yield_draws, rates, price_grain = 300, price_input = 1.2,
                  grounding = grounding_grounded())
#> <decision> decided  [grounded]
#>   action : rate=175
#>   E[utility]: 1356
#>   candidates: 9 evaluated (method: expected_profit)

# A trading decision: size a position by drawdown-capped expected log-growth.
# An un-grounded price feed forces flat -- the capital firewall.
returns <- rnorm(5000L, 0.02, 0.03)
decide_position_size(returns, capital = 1000, max_drawdown = 0.15,
                     grounding = grounding_unverified())
#> <decision> ABSTAINED (input_ungrounded)  [ungrounded]
#>   action : flat (0%) [ungrounded]
#>   candidates: 41 evaluated (method: kelly_log_growth_drawdown_capped)

# A quality-graded payoff: protein bands price the decision, not the mean.
protein_value <- grade_band_value(breaks = c(11.5, 13.0),
                                  values = c(300, 360, 420))
protein_value
#> <grade_band_value> 3 bands
#>          band value
#>  [-Inf, 11.5)   300
#>    [11.5, 13)   360
#>     [13, Inf)   420

See vignette("getting-started") for the full walk-through, including the manifest-native pipeline tail and the grade-band loss worked end to end.

Contributing

Bug reports and suggestions are welcome through the issue tracker.

Citation

citation("decideR")

License

This package is released under: MIT + file LICENSE.