Domain
|
Description
|
Problem Generators
|
Problem Sets
|
|
Gold-Miner
|
A robot is in a mine and has the goal of
reaching a location that contains gold. The mine is organized as a grid
with each cell either begin hard or soft rock. There is a special
location where the robot can either pickup an endlesssupply of bombs or
pickup a laser cannon. The laser cannon can shoot through both hard and
soft rock, whereas the bomb can only penetrate soft rock. However, the
laser cannon also will destroy the gold if used to uncover the gold
location. The bomb will not destroy the gold. The problem difficulty is
scaled by increasing the size of the grid.
This domain has a simple optimal strategy:
1) get the laser cannon,
2) shoot through the rock until reaching a cell bordering the gold.
3) go and get a bomb.
4) blast away the rock at the gold location.
5) pickup the gold.
|
|
Learning
Phase Distribution
Testing
Phase Distribution
|
|
|
Matching Blocksworld
|
This is a simple variant of the blocks
world where each block is either positive or negative and there are two
hands, one positive and one negative. The twist is that if a block is
picked up by a hand of opposite polarity then it is damaged such that
no other block can be placed on it, which can lead to deadends. The
interaction between hands and blocks of the same polarity is just as in
the standard blocks world.
|
|
Learning
Phase Distribution
Testing
Phase Distribution
|
|
|
N-Puzzle
|
This is the classic NxN sliding puzzle
domain. This has been a common testbed for search algorithms and in
particular work on macro learning.
|
|
Learning
Phase Distribution
Testing
Phase Distribution
|
|
|
Parking
|
This domain involves parking cars on a
street with N curb locations where cars can be double parked but not
triple parked. The goal is to move from one configuration of parked
cars to another configuration by driving cars from one curb location to
another. The problems in the competition contain 2*(N-1) cars, which
allows one free curb space and guarantees solvability. The problem
difficulty is scaled by increasing the number of curb locations. This
domain is isomorphic to a bounded-table blocks world domain where there
are N table locations and 2*(N-1) blocks and towers are only allowed to
contain 2 blocks.
|
|
Learning
Phase Distribution
Testing
Phase Distribution
|
|
|
Sokoban
|
This domain is inspired by the popular
Sokoban puzzle game where an agent has the goal of pushing a set of
boxes into specified goal locations in a grid with walls. The
competition problems are generated in a way that guarantees solvability
and generally are easy problem instances for humans.
|
|
Learning
Phase Distribution
Testing
Phase Distribution
|
|
|
Thoughtful
|
This domain models a simplified version of
Thoughtful Solitaire, which is a solitaire variant that is played with
all of the cards visible. The rules follow those described in the below
paper but simplified so that one can turn each card from the talon
rather than 3 cards at a time. The problem difficulty is scaled by
varying the number of cards and stacks. The competition problems have
all been proven to be solvable using the solver described in the paper
below.
Searching
Solitaire in Real Time. (2007). Ronald Bjarnason, Prasad Tadepalli,
and Alan Fern. International Computer Games Association Journal, 30
(3), 131-142.S
|
|
Learning
Phase Distribution
Testing
Phase Distribution
|
|
|
|
|
All
Learning Phase Distributions
All
Testing Phase Distributions
|