Cliff_walking: Cliff Walking Gridworld MDP

In pomdp: Infrastructure for Partially Observable Markov Decision Processes (POMDP)

Cliff Walking Gridworld MDP

Description

The cliff walking gridworld MDP example from Chapter 6 of the textbook "Reinforcement Learning: An Introduction."

Format

An object of class MDP.

Details

The cliff walking gridworld has the following layout:

The gridworld is represented as a 4 x 12 matrix of states. The states are labeled with their x and y coordinates. The start state is in the bottom left corner. Each action has a reward of -1, falling off the cliff has a reward of -100 and returns the agent back to the start. The episode is finished once the agent reaches the absorbing goal state in the bottom right corner. No discounting is used (i.e., \gamma = 1).

References

Richard S. Sutton and Andrew G. Barto (2018). Reinforcement Learning: An Introduction Second Edition, MIT Press, Cambridge, MA.

See Also

Other MDP_examples: MDP(), Maze, Windy_gridworld

Other gridworld: Maze, Windy_gridworld, gridworld

Examples

data(Cliff_walking)
Cliff_walking

gridworld_matrix(Cliff_walking)
gridworld_matrix(Cliff_walking, what = "labels")

# The Goal is an absorbing state 
which(absorbing_states(Cliff_walking))

# visualize the transition graph
gridworld_plot_transition_graph(Cliff_walking)

# solve using different methods
sol <- solve_MDP(Cliff_walking) 
sol
policy(sol)
gridworld_plot_policy(sol)

sol <- solve_MDP(Cliff_walking, method = "q_learning", N = 100) 
sol
policy(sol)
gridworld_plot_policy(sol)

sol <- solve_MDP(Cliff_walking, method = "sarsa", N = 100) 
sol
policy(sol)
gridworld_plot_policy(sol)

sol <- solve_MDP(Cliff_walking, method = "expected_sarsa", N = 100, alpha = 1) 
policy(sol)
gridworld_plot_policy(sol)

pomdp documentation built on May 29, 2024, 2:04 a.m.