Skip to content

Latest commit

 

History

History
66 lines (41 loc) · 3.53 KB

README.md

File metadata and controls

66 lines (41 loc) · 3.53 KB

cpp/ structure

algorithms/

Provides definitions for all the algorithms (AURORA, TAXONS, and their Hand-Coded variants).

Each algorithm has a modifier type (modifier_t), and a stat type (stat_t), both following the sferes2 framework. Moreover, each algorithm defines a selector type (selector_t), and a container type (container_t), following the Quality Diversity framework See for example the definitions of AURORA variants.

environments/

Source code for all tasks under study (Maze, Hexapod and Air-Hockey).

Each environment has a fitness type (fit_t), genotype type (gen_t), and phenotype type (phen_t), all following the sferes2 framework.

modifier/

The modifier is executed after having mutated, copied and evaluated the population.

Modifier for AURORA and TAXONS

In the case of AURORA and TAXONS, we use the modifier DimensionalityReduction, which has several roles:

  • Using the Encoder to encode the sensory data from new individuals into Behavioural Descriptors (BDs).
  • Encoder Update: Training the Encoder based on all the sensory data from the container, and then use it to recompute all the BDs.
  • Container Size Control (CSC): Updating the value of the distance threshold (d_{min}).
  • Container Update: Performing the container update by removing all individuals from the container, and re-adding them with their recomputed BDs, and with the up-to-date distance threshold d_{min}.

This modifier uses a NetworkLoader as an interface with the torch models. For example, the NetworkLoader can be called to evaluate the latent projection of a model, and to train the torch model on a dataset.

The torch models are stored in the autoencoder folder.

Modifier for hand-coded variants

Hand-coded variants use a different modifier ContainerUpdateHandCoded, which does not consider any dimensionality reduction algorithm. In other words, the hand-coded variants only perform the last two items described above:

  • Container Size Control (CSC): Updating the value of the distance threshold (d_{min}).
  • Container Update: Performing the container update by removing all individuals from the container, and re-adding them with their recomputed BDs, and with the up-to-date distance threshold d_{min}.

stat/

The Statistics classes are used to save the relevant data for further analysis.

The most useful Statistics class is Projection which periodically saves the following attributes of all individuals from the archive:

  • index of the individual
  • reconstruction error of its sensory data by the AE
  • Unsupervised behavioural descriptor
  • Hand-coded behavioural descriptor (ground-truth state)
  • Fitness value (equal to -1 if the fitness function is not considered)
  • "Implicit" Fitness value (present the fitness value that would be obtained if the fitness function was considered)
  • Novelty score

Source Files

  • aurora.cpp - for defining main experiments.

  • save_video.cpp - used to easily save videos for the Hexapod and Air-Hockey tasks.

  • train_ae_from_scratch.cpp - used to re-train the AE from scratch given a pre-defined dataset (has been mostly used for the Maze task).

Others

  • compilation_variables.hpp - making the link between the variables defined in the Experiment object, and the C++ compilation variables.