HDOSE description 🔗

The High Dimension Origin Space Exploration (HDOSE) method is used to explore the multiple antecedents of a pattern when the input space has high dimensionality (i.e., > 6 dimensions). It generates input parameter values that produce a pattern described by a set of constraints on the objectives (i.e., objectives should be either below given thresholds or close to specified target values). HDOSE optimizes the fitness and archives the solutions that are good enough. New solutions are added to the archive when their input values are at least distance d = 1.0 apart from all existing solutions in the archive (using L1 distance). When the archive size reaches 500, a new value for d is computed to reduce the archive size to below 500, ensuring that all solutions in the archive remain at least distance d apart. The hook arguments for the HDOSEEvolution are:

• output: the file in which to store the results,
• keepHistory: optional, Boolean, keep the history of the results for future analysis,
• frequency: optional, Long, the frequency in generations where the result should be saved, it is generally set to avoid using too much disk space,
• keepAll: optional, Boolean, save all the individuals of the population not only the optimal ones.

Example 🔗

This is an example of the HDOSE method in an OpenMOLE script:

val myseed = Val[Long]

val param1 = Val[Array[Double]]
val param2 = Val[Double]
val param3 = Val[Array[Boolean]]
val param4 = Val[Array[Int]]

val output1 = Val[Double]
val output2 = Val[Double]

HDOSEEvolution(
  evaluation = modelTask,
  parallelism = 10,
  termination = 100,
  origin = Seq(
    param1 in Seq.fill(10)(0.0 to 1.0),
    param2 in (-10.0 to 10.0 weight 10.0),
    param3 in Seq.fill(5)(TrueFalse weight 2.0),
    param4 in Seq.fill(10)(0 to 100)
  ),
  objective = Seq(
    output1 under 5.0,
    output2 under 50.0),
  stochastic = Stochastic(seed = myseed)
) hook (workDirectory / "results.omr", frequency = 100)

origin describes the genome of the optimisation. To compute the distance, each part of the genome is normalised. To give more weight to a given part of the genome in the distance computation, use the `weight` keyword.
objective uses inequalities to define the pattern to reach. The pattern is reached when all objectives are below their threshold value. In this example, HDOSE finds the maximum diversity of inputs for which all outputs are below their respective thresholds.