Glossary - Design of Experiments for Reinforcement Learning - page 182

Glossary

Action selection policy The method by which the agent selects an ac-

tion. This work uses an -greedy action selection
policy.

Agent An entity that interacts with the environment or

domain. The agent is comprised of a functional
representation of its knowledge, a method of
learning and updating its knowledge about the
environment, and an action selection policy.

Benchmark problems Simple problems used by the reinforcement

learning community to test, evaluate, and under-
stand new algorithms and learning approaches.
Some common benchmark problems include
Gridworld, the mountain car problem, cart-pole
balancing, and the pendulum swing-up problem.

CART (Classification And Regression Trees) A classifi-

cation or regression model that results in a binary
tree that partitions the input parameter space into
non-overlapping hyperrectangular regions. See
Breiman (2001).

Covariance kernel Kernel function which accounts for the spatial

variance between points in a kriging model.

Design of experiments The formal process of designing an experimental

protocol and analyzing the empirically collected
data in order to discover valid and objective
information about an underlying system. See
Montgomery (2008).

Empirical convergence A region of the learning performance curve

where learning stabilizes. This work uses a mov-
ing proportion of how often the goal is reached as
a performance metric. Convergence is reached if
the level, range, and slope of this curve passes
threshold values.