This contains a set of MDPs to test your code on. They are all fairly simple, so reading the code should help explain most of it.
The only functions needed for a policy learning/q-learning agent are
given in mdp.py. There are three different domains:
-
Discrete Grid World in
gridworld.py: An agent can move up, down, left and right with some noise, and must avoid pits to reach some goal. -
Continuous Grid World in
gridworld_c.py: Continuous variant of the above. -
Wumpus domain: Similar to discrete grid world, but wumpuses kill and pits only cause some damage. The agent can pick up gold for extra reward or immunity to pass through wumpuses.
There are also some sample simulations of each domain given in the
-sample.py files.