Planning under uncertainty

Introduction

• Decisions under uncertainty.

Planning Under Uncertainty MDP

MDP = Markov Decision Processes POMDP = Partially Observable Markov Decision Processes

Robot Tour Guide Examples

MDP Grid World

Problems with Conventional Planning 1

Branching Factor is too large

Quiz: Branching Factor Question

For this problem (and only this problem) assume actions are stochastic in a way that is different than described in 4. MDP Gridworld.

Instead of an action north possibly going east or west, an action north will possibly go northeast or northwest (i.e. to the diagonal squares).

Likewise for the other directions e.g. an action west will possibly go west, northwest or southwest (i.e. to the diagonals).

Problems With Conventional Planning 2

Quiz: Policy Question 1

Stochastic actions are as in 4. MDP Grid World.

An action North moves North with 80% chance otherwise East with 10% chance or West with 10% chance. Likewise for the other directions.

Quiz: Policy Question 2

Stochastic actions are as in 4. MDP Grid World.

An action North moves North with 80% chance otherwise East with 10% chance or West with 10% chance. Likewise for the other directions.

Quiz: Policy Question 3

Stochastic actions are as in 4. MDP Grid World.

An action North moves North with 80% chance otherwise East with 10% chance or West with 10% chance. Likewise for the other directions.

MDP And Costs

Value Iteration 1

Value Iteration 2

Value Iteration 3

Quiz: Deterministic Question 1

Quiz: Deterministic Question 2

Quiz: Deterministic Question 3

Quiz: Stochastic Question 1

Quiz: Stochastic Question 2

>>> 77 * 0.8 + (0.1 * -100) - 3
48.6

Value Iterations and Policy 1

Value Iterations And Policy 2

Markov Decision Process Conclusion

Partial Observability Introduction

• http://robots.stanford.edu/papers/EmeryMontemerlo04a.pdf

POMDP Vs MDP

POMDP

Planning Under Uncertainity Conclusion

Further Study

Charles Isbell and Michael Littmann’s ML course

• Markov Decision Processes

• Reinforcement Learning

Peter Norvig and Sebastian Thrun’s AI course:

• Reinforcement Learning_2