Skip to content

Scenarios overview

juremaj edited this page Dec 22, 2022 · 3 revisions

Complexity scenarios

Apple and Gabor gathering

These are here mostly for puproses of archiving/recreating old results. The naming conventions are also somewhat arbitrary. In the apples task r refers to the number of red apples initially spawned, b refers to blue apples and g refers to grass. hr100 refers to the reward being defined as current health multiplied by 100. The numbers in the suffices are just internal references due to changing minor things or debugging.

The assignments to health values are as follows:

NOTE: The exact health contributions might be different, due to some balancing across tasks. So to be sure check the SCRIPTS in the .wad file! (The ordering of the stimuli/categories is still the same though)

Apple gathering:
Picture1

Gabor gathering:
assignments_01

MNIST gathering

The assignments for MNIST gathering are the following:
assignments_1
Here the particular images given are just example of a category, when spawning stimuli a random image from that category is chosen.

CIFAR gathering

Here the 01 version has a one-to-one correspondence with MNIST 01 in terms of game mechanics.

The assignments for CIFAR gathering are the following: image

Distractor scenarios

APPCIFAR gathering

Here the scenarios always come in pairs of appcifar_apples_gathering_... and appcifar_cifar_gathering_..., where the set of stimuli referenced by the title represents the valent ones. For example in appcifar_apples_gathering_... the agent's health is modulated only by apples, collecting CIFAR stimuli has no effect on its health (vice versa for appcifar_cifar_gathering_...). Important: The ..._06 scenarios have rewards for all positive and all negative stimuli lumped into two categories. It is also approximately balanced with the other scenarios in terms of cumulative spawned reward, in order to be able to compare the performance better across scenarios.

APPMNIST gathering

This is the same as above, just using MNIST instead of CIFAR. This might be better, since apples need color separation and MNIST needs shape separation. In the APPCIFAR these are both mixed together.

Obstacle scenarios

These are all the same as the original MNIST scenario, the only addition is grass obstacles. There are three versions with 100, 200 and 300 obstacles respectively.