An Introduction to Reinforcement Learning

Shivaram Kalyanakrishnan
[email protected]

Department of Computer Science and Automation

Indian Institute of Science

August 2014
What is Reinforcement Learning?

Shivaram Kalyanakrishnan 1/19

What is Reinforcement Learning?
[Video1 of little girl learning to ride bicycle]

1. https://www.youtube.com/watch?v=Qv43pKlVZXk

Shivaram Kalyanakrishnan 1/19

What is Reinforcement Learning?
[Video1 of little girl learning to ride bicycle]

1. https://www.youtube.com/watch?v=Qv43pKlVZXk

Shivaram Kalyanakrishnan 1/19

What is Reinforcement Learning?
[Video1 of little girl learning to ride bicycle]

Learning to Drive a Bicycle using Reinforcement Learning and Shaping

Jette Randløv and Preben Alstrøm. ICML 1998.

1. https://www.youtube.com/watch?v=Qv43pKlVZXk

Shivaram Kalyanakrishnan 1/19

What is Reinforcement Learning?
[Video1 of little girl learning to ride bicycle]

Learning to Drive a Bicycle using Reinforcement Learning and Shaping

Jette Randløv and Preben Alstrøm. ICML 1998.

Learning by trial and error to perform sequential decision making.

1. https://www.youtube.com/watch?v=Qv43pKlVZXk

Shivaram Kalyanakrishnan 1/19

Our View of Reinforcement Learning

Operations Research
Control Theory
(Dynamic Programming)

Psychology Reinforcement Neuroscience

(Animal Behaviour) Learning

Artificial Intelligence and

Computer Science

Shivaram Kalyanakrishnan 2/19

Our View of Reinforcement Learning

Operations Research
Control Theory
(Dynamic Programming)

Psychology Reinforcement Neuroscience

(Animal Behaviour) Learning

Artificial Intelligence and

Computer Science

B. F. Skinner

Shivaram Kalyanakrishnan 2/19

Our View of Reinforcement Learning

Operations Research
Control Theory
(Dynamic Programming)
R. E. Bellman

Psychology Reinforcement Neuroscience

(Animal Behaviour) Learning

Artificial Intelligence and

Computer Science

B. F. Skinner

Shivaram Kalyanakrishnan 2/19

Our View of Reinforcement Learning

Operations Research
Control Theory
(Dynamic Programming)
R. E. Bellman D. P. Bertsekas

Psychology Reinforcement Neuroscience

(Animal Behaviour) Learning

Artificial Intelligence and

Computer Science

B. F. Skinner

Shivaram Kalyanakrishnan 2/19

Our View of Reinforcement Learning

Operations Research
Control Theory
(Dynamic Programming)
R. E. Bellman D. P. Bertsekas

Psychology Reinforcement Neuroscience

(Animal Behaviour) Learning

Artificial Intelligence and

Computer Science

B. F. Skinner W. Schultz

Shivaram Kalyanakrishnan 2/19

Our View of Reinforcement Learning

Operations Research
Control Theory
(Dynamic Programming)
R. E. Bellman D. P. Bertsekas

Psychology Reinforcement Neuroscience

(Animal Behaviour) Learning

Artificial Intelligence and

Computer Science

B. F. Skinner W. Schultz

R. S. Sutton

Shivaram Kalyanakrishnan 2/19

Resources
Reinforcement Learning: A Survey.
Leslie Pack Kaelbling, Michael L. Littman, and Andrew W. Moore. JAIR 1996.

Shivaram Kalyanakrishnan 3/19

Resources
Reinforcement Learning: A Survey.
Leslie Pack Kaelbling, Michael L. Littman, and Andrew W. Moore. JAIR 1996.

Reinforcement Learning: An Introduction

Richard S. Sutton and Andrew G. Barto. MIT Press, 1998.

Algorithms for Reinforcement Learning

Csaba Szepesvári. Morgan & Claypool, 2010.

Shivaram Kalyanakrishnan 3/19

Resources
Reinforcement Learning: A Survey.
Leslie Pack Kaelbling, Michael L. Littman, and Andrew W. Moore. JAIR 1996.

Reinforcement Learning: An Introduction

Richard S. Sutton and Andrew G. Barto. MIT Press, 1998.

Algorithms for Reinforcement Learning

Csaba Szepesvári. Morgan & Claypool, 2010.

E-mail List: [email protected].

Shivaram Kalyanakrishnan 3/19

Resources
Reinforcement Learning: A Survey.
Leslie Pack Kaelbling, Michael L. Littman, and Andrew W. Moore. JAIR 1996.

Reinforcement Learning: An Introduction

Richard S. Sutton and Andrew G. Barto. MIT Press, 1998.

Algorithms for Reinforcement Learning

Csaba Szepesvári. Morgan & Claypool, 2010.

E-mail List: [email protected].

RL Competition: http://www.rl-competition.org/.

Shivaram Kalyanakrishnan 3/19

Today’s Class

1. Markov decision problems

2. Bellman’s (Optimality) Equations, planning and learning
3. Challenges
4. Summary

Shivaram Kalyanakrishnan 4/19

Today’s Class

1. Markov decision problems

2. Bellman’s (Optimality) Equations, planning and learning
3. Challenges
4. Summary

Shivaram Kalyanakrishnan 4/19

Markov Decision Problem
ENVIRONMENT
st+1
T

R
r t+1

state reward action

st rt at
LEARNING AGENT

π: S A

S: set of states.
A: set of actions.
T : transition function. ∀s ∈ S, ∀a ∈ A, T (s, a) is a distribution over S.
R: reward function. ∀s, s′ ∈ S, ∀a ∈ A, R(s, a, s′ ) is a finite real number.
γ: discount factor. 0 ≤ γ < 1.

Shivaram Kalyanakrishnan 5/19

Markov Decision Problem
ENVIRONMENT
st+1
T

R
r t+1

state reward action

st rt at
LEARNING AGENT

π: S A

S: set of states.
A: set of actions.
T : transition function. ∀s ∈ S, ∀a ∈ A, T (s, a) is a distribution over S.
R: reward function. ∀s, s′ ∈ S, ∀a ∈ A, R(s, a, s′ ) is a finite real number.
γ: discount factor. 0 ≤ γ < 1.

Trajectory over time: s0 , a0 , r1 , s1 , a1 , r2 , . . . , st , at , rt+1 , st+1 , . . . .

Shivaram Kalyanakrishnan 5/19

Markov Decision Problem
ENVIRONMENT
st+1
T

R
r t+1

state reward action

st rt at
LEARNING AGENT

π: S A

S: set of states.
A: set of actions.
T : transition function. ∀s ∈ S, ∀a ∈ A, T (s, a) is a distribution over S.
R: reward function. ∀s, s′ ∈ S, ∀a ∈ A, R(s, a, s′ ) is a finite real number.
γ: discount factor. 0 ≤ γ < 1.

Trajectory over time: s0 , a0 , r1 , s1 , a1 , r2 , . . . , st , at , rt+1 , st+1 , . . . .

Value, or expected long-term reward, of state s under policy π:
V π (s) = E[r1 + γr2 + γ 2 r3 + . . . to ∞|s0 = s, ai = π(si )].

Shivaram Kalyanakrishnan 5/19

Markov Decision Problem
ENVIRONMENT
st+1
T

R
r t+1

state reward action

st rt at
LEARNING AGENT

π: S A

S: set of states.
A: set of actions.
T : transition function. ∀s ∈ S, ∀a ∈ A, T (s, a) is a distribution over S.
R: reward function. ∀s, s′ ∈ S, ∀a ∈ A, R(s, a, s′ ) is a finite real number.
γ: discount factor. 0 ≤ γ < 1.

Trajectory over time: s0 , a0 , r1 , s1 , a1 , r2 , . . . , st , at , rt+1 , st+1 , . . . .

Value, or expected long-term reward, of state s under policy π:
V π (s) = E[r1 + γr2 + γ 2 r3 + . . . to ∞|s0 = s, ai = π(si )].

Objective: “Find π such that V π (s) is maximal ∀s ∈ S.”

Shivaram Kalyanakrishnan 5/19

Examples
What are the agent and environment? What are S, A, T , and R?

Shivaram Kalyanakrishnan 6/19

Examples
What are the agent and environment? What are S, A, T , and R?

1. http://www.chess-game-strategies.com/images/kqa_chessboard_large-picture_2d.gif

Shivaram Kalyanakrishnan 6/19

Examples
What are the agent and environment? What are S, A, T , and R?

An Application of Reinforcement Learning to Aerobatic Helicopter Flight

Pieter Abbeel, Adam Coates, Morgan Quigley, and Andrew Y. Ng. NIPS 2006.

1. http://www.chess-game-strategies.com/images/kqa_chessboard_large-picture_2d.gif
2. http://www.aviationspectator.com/files/images/
SH-3-Sea-King-helicopter-191.preview.jpg

Shivaram Kalyanakrishnan 6/19

Examples
What are the agent and environment? What are S, A, T , and R?

[Video3 of Tetris]

An Application of Reinforcement Learning to Aerobatic Helicopter Flight

Pieter Abbeel, Adam Coates, Morgan Quigley, and Andrew Y. Ng. NIPS 2006.

1. http://www.chess-game-strategies.com/images/kqa_chessboard_large-picture_2d.gif
2. http://www.aviationspectator.com/files/images/
SH-3-Sea-King-helicopter-191.preview.jpg
3. https://www.youtube.com/watch?v=khHZyghXseE

Shivaram Kalyanakrishnan 6/19

Illustration: MDPs as State Transition Diagrams
0.2, 0

0.8, 1
s1 s2 0.2, 1
0.5, 1
0.5, −1

0.8, 0 0.5, 0 0.5, 2 0.8, −1

0.5, 1 0.5, −1
0.5, 0
0.5, 2
0.2, 2 s4 s3
0.8, 2

0.2, −1
Notation: "transition probability, reward" marked on each arrow

States: s1 , s2 , s3 , and s4 .
Actions: Red (solid lines) and blue (dotted lines).
Transitions: Red action leads to same state with 20% chance, to next-clockwise state with
80% chance. Blue action leads to next-clockwise state or 2-removed-clockwise state with
equal (50%) probability.
Rewards: R(∗, ∗, s1 ) = 0, R(∗, ∗, s2 ) = 1, R(∗, ∗, s3 ) = −1, R(∗, ∗, s4 ) = 2.
Discount factor: γ = 0.9.

Shivaram Kalyanakrishnan 7/19

Today’s Class

1. Markov decision problems

2. Bellman’s (Optimality) Equations, planning and learning
3. Challenges
4. Summary

Shivaram Kalyanakrishnan 8/19

Bellman’s Equations
Recall that
V π (s) = E[r1 + γr2 + γ 2 r3 + . . . |s0 = s, ai = π(si )].
Bellman’s Equations (∀s ∈ S):
P
V π (s) = s′ ∈S T (s, π(s), s′ ) [R(s, π(s), s′ ) + γV π (s′ )].
V π is called the value function of π.

Shivaram Kalyanakrishnan 9/19

Bellman’s Equations
Recall that
V π (s) = E[r1 + γr2 + γ 2 r3 + . . . |s0 = s, ai = π(si )].
Bellman’s Equations (∀s ∈ S):
P
V π (s) = s′ ∈S T (s, π(s), s′ ) [R(s, π(s), s′ ) + γV π (s′ )].
V π is called the value function of π.

Define (∀s ∈ S, ∀a ∈ A):

P
Q π (s, a) = s′ ∈S T (s, a, s′ ) [R(s, a, s′ ) + γV π (s′ )].
Q π is called the action value function of π.
V π (s) = Q π (s, π(s)).

Shivaram Kalyanakrishnan 9/19

Bellman’s Equations
Recall that
V π (s) = E[r1 + γr2 + γ 2 r3 + . . . |s0 = s, ai = π(si )].
Bellman’s Equations (∀s ∈ S):
P
V π (s) = s′ ∈S T (s, π(s), s′ ) [R(s, π(s), s′ ) + γV π (s′ )].
V π is called the value function of π.

Define (∀s ∈ S, ∀a ∈ A):

P
Q π (s, a) = s′ ∈S T (s, a, s′ ) [R(s, a, s′ ) + γV π (s′ )].
Q π is called the action value function of π.
V π (s) = Q π (s, π(s)).

The variables in Bellman’s equation are the V π (s). |S| linear equations
in |S| unknowns.

Shivaram Kalyanakrishnan 9/19

Bellman’s Equations
Recall that
V π (s) = E[r1 + γr2 + γ 2 r3 + . . . |s0 = s, ai = π(si )].
Bellman’s Equations (∀s ∈ S):
P
V π (s) = s′ ∈S T (s, π(s), s′ ) [R(s, π(s), s′ ) + γV π (s′ )].
V π is called the value function of π.

Define (∀s ∈ S, ∀a ∈ A):

P
Q π (s, a) = s′ ∈S T (s, a, s′ ) [R(s, a, s′ ) + γV π (s′ )].
Q π is called the action value function of π.
V π (s) = Q π (s, π(s)).

The variables in Bellman’s equation are the V π (s). |S| linear equations
in |S| unknowns.

Thus, given S, A, T , R, γ, and a fixed policy π, we can solve Bellman’s

equations efficiently to obtain, ∀s ∈ S, ∀a ∈ A, V π (s) and Q π (s, a).

Shivaram Kalyanakrishnan 9/19

Bellman’s Optimality Equations
Let Π be the set of all policies. What is its cardinality?

Shivaram Kalyanakrishnan 10/19

Bellman’s Optimality Equations
Let Π be the set of all policies. What is its cardinality?

It can be shown that there exists a policy π ∗ ∈ Π such that

∗
∀π ∈ Π ∀s ∈ S: V π (s) ≥ V π (s).
∗ ∗
V π is denoted V ∗ , and Q π is denoted Q ∗ .
There could be multiple optimal policies π ∗ , but V ∗ and Q ∗ are unique.

Shivaram Kalyanakrishnan 10/19

Bellman’s Optimality Equations
Let Π be the set of all policies. What is its cardinality?

It can be shown that there exists a policy π ∗ ∈ Π such that

∗
∀π ∈ Π ∀s ∈ S: V π (s) ≥ V π (s).
∗ ∗
V π is denoted V ∗ , and Q π is denoted Q ∗ .
There could be multiple optimal policies π ∗ , but V ∗ and Q ∗ are unique.

Bellman’s Optimality Equations (∀s ∈ S):

P
V ∗ (s) = maxa∈A s′ ∈S T (s, a, s′ ) [R(s, a, s′ ) + γV ∗ (s′ )].

Shivaram Kalyanakrishnan 10/19

Bellman’s Optimality Equations
Let Π be the set of all policies. What is its cardinality?

It can be shown that there exists a policy π ∗ ∈ Π such that

∗
∀π ∈ Π ∀s ∈ S: V π (s) ≥ V π (s).
∗ ∗
V π is denoted V ∗ , and Q π is denoted Q ∗ .
There could be multiple optimal policies π ∗ , but V ∗ and Q ∗ are unique.

Bellman’s Optimality Equations (∀s ∈ S):

P
V ∗ (s) = maxa∈A s′ ∈S T (s, a, s′ ) [R(s, a, s′ ) + γV ∗ (s′ )].

Planning problem:
Given S, A, T , R, γ, how can we find an optimal policy π ∗ ? We need
to be computationally efficient.

Shivaram Kalyanakrishnan 10/19

Bellman’s Optimality Equations
Let Π be the set of all policies. What is its cardinality?

It can be shown that there exists a policy π ∗ ∈ Π such that

∗
∀π ∈ Π ∀s ∈ S: V π (s) ≥ V π (s).
∗ ∗
V π is denoted V ∗ , and Q π is denoted Q ∗ .
There could be multiple optimal policies π ∗ , but V ∗ and Q ∗ are unique.

Bellman’s Optimality Equations (∀s ∈ S):

P
V ∗ (s) = maxa∈A s′ ∈S T (s, a, s′ ) [R(s, a, s′ ) + γV ∗ (s′ )].

Planning problem:
Given S, A, T , R, γ, how can we find an optimal policy π ∗ ? We need
to be computationally efficient.

Learning problem:
Given S, A, γ, and the facility to follow a trajectory by sampling from T
and R, how can we find an optimal policy π ∗ ? We need to be sample-
efficient.

Shivaram Kalyanakrishnan 10/19

Planning
Given S, A, T , R, γ, how can we find an optimal policy π ∗ ?

Shivaram Kalyanakrishnan 11/19

Planning
Given S, A, T , R, γ, how can we find an optimal policy π ∗ ?

One way. We can pose Bellman’s optimality equations as a linear

program, solve for V ∗ , derive Q ∗ , and induce π ∗ (s) = argmaxa Q ∗ (s, a).

Shivaram Kalyanakrishnan 11/19

Planning
Given S, A, T , R, γ, how can we find an optimal policy π ∗ ?

One way. We can pose Bellman’s optimality equations as a linear

program, solve for V ∗ , derive Q ∗ , and induce π ∗ (s) = argmaxa Q ∗ (s, a).
Another way. We can apply the policy iteration algorithm, which is
typically more efficient in practice.

Pick an initial policy π arbitrarily.

Compute Q π using Bellman’s equations.
converged ← false.
Repeat
Set π ′ as: ∀s ∈ S, π ′ (s) = argmaxa Q π (s, a)
(break ties arbitrarily).[Improvement]
′
Compute Q π using Bellman’s equations. [Evaluation]
′
If (Q π = Q π ), converged ← true.
′
π ← π ′ , Q π ← Q π .
Until converged.
Return π (which is provably optimal).

Shivaram Kalyanakrishnan 11/19

Planning
Given S, A, T , R, γ, how can we find an optimal policy π ∗ ?

One way. We can pose Bellman’s optimality equations as a linear

program, solve for V ∗ , derive Q ∗ , and induce π ∗ (s) = argmaxa Q ∗ (s, a).
Another way. We can apply the policy iteration algorithm, which is
typically more efficient in practice.

Pick an initial policy π arbitrarily.

Compute Q π using Bellman’s equations.
converged ← false.
Repeat
Set π ′ as: ∀s ∈ S, π ′ (s) = argmaxa Q π (s, a)
(break ties arbitrarily).[Improvement]
′
Compute Q π using Bellman’s equations. [Evaluation]
′
If (Q π = Q π ), converged ← true.
′
π ← π ′ , Q π ← Q π .
Until converged.
Return π (which is provably optimal).

Other ways. Value iteration and its various “mixtures” with policy
iteration.
Shivaram Kalyanakrishnan 11/19
Learning
Given S, A, γ, and the facility to follow a trajectory by sampling from T
and R, how can we find an optimal policy π ∗ ?

Shivaram Kalyanakrishnan 12/19

Learning
Given S, A, γ, and the facility to follow a trajectory by sampling from T
and R, how can we find an optimal policy π ∗ ?

Various classes of learning methods exist. We will consider a simple one

called Q-learning, which is a temporal difference learning algorithm.

 Let Q be our “guess” of Q ∗ : for every state s and action a, initialise

Q(s, a) arbitrarily. We will start in some state s0 .
For t = 0, 1, 2, . . .
Take an action at , chosen uniformly at random with probability ǫ,
and to be argmaxa Q(st , a) with probability 1 − ǫ.
The environment will generate next state st+1 and reward rt+1 .
Update: Q(st , at ) ← Q(st , at ) + αt (rt +1 + γ maxa∈A Q(st +1 , a) − Q(st , at )).
[ǫ: parameter for “ǫ-greedy” exploration] [αt : learning rate]
[rt +1 + γ maxa∈A Q(st +1 , a) − Q(st , at ): temporal difference prediction error]

Shivaram Kalyanakrishnan 12/19

Learning
Given S, A, γ, and the facility to follow a trajectory by sampling from T
and R, how can we find an optimal policy π ∗ ?

Various classes of learning methods exist. We will consider a simple one

called Q-learning, which is a temporal difference learning algorithm.

 Let Q be our “guess” of Q ∗ : for every state s and action a, initialise

Q(s, a) arbitrarily. We will start in some state s0 .
For t = 0, 1, 2, . . .
Take an action at , chosen uniformly at random with probability ǫ,
and to be argmaxa Q(st , a) with probability 1 − ǫ.
The environment will generate next state st+1 and reward rt+1 .
Update: Q(st , at ) ← Q(st , at ) + αt (rt +1 + γ maxa∈A Q(st +1 , a) − Q(st , at )).
[ǫ: parameter for “ǫ-greedy” exploration] [αt : learning rate]
[rt +1 + γ maxa∈A Q(st +1 , a) − Q(st , at ): temporal difference prediction error]

For ǫ ∈ (0, 1] and αt = 1t , it can be proven that as t → ∞, Q → Q ∗ .

Q-Learning
Christopher J. C. H. Watkins and Peter Dayan. Machine Learning, 1992.

Shivaram Kalyanakrishnan 12/19

Today’s Class

1. Markov decision problems

2. Bellman’s (Optimality) Equations, planning and learning
3. Challenges
4. Summary

Shivaram Kalyanakrishnan 13/19

Practical Implementation and Evaluation of Learning Algorithms

Shivaram Kalyanakrishnan 14/19

Practical Implementation and Evaluation of Learning Algorithms
Generalized Model Learning for Reinforcement Learning on a Humanoid Robot
Todd Hester, Michael Quinlan, and Peter Stone. ICRA 2010.

[Video1 of RL on a humanoid robot]

1. http://www.youtube.com/watch?v=mRpX9DFCdwI

Shivaram Kalyanakrishnan 14/19

Practical Implementation and Evaluation of Learning Algorithms
Generalized Model Learning for Reinforcement Learning on a Humanoid Robot
Todd Hester, Michael Quinlan, and Peter Stone. ICRA 2010.

[Video1 of RL on a humanoid robot]

1. http://www.youtube.com/watch?v=mRpX9DFCdwI

Shivaram Kalyanakrishnan 14/19

Research Challenges

 Exploration
 Generalisation (over states and actions)
 State aliasing (partial observability)
 Multiple agents, nonstationary rewards and transitions
 Abstraction (over states and over time)
 Proofs of convergence, sample-complexity bounds

Shivaram Kalyanakrishnan 15/19

Research Challenges

 Exploration
 Generalisation (over states and actions)
 State aliasing (partial observability)
 Multiple agents, nonstationary rewards and transitions
 Abstraction (over states and over time)
 Proofs of convergence, sample-complexity bounds

My thesis question:

“How well do different learning methods for sequential decision making

perform in the presence of state aliasing and generalization; can we de-
velop methods that are both sample-efficient and capable of achieving
high asymptotic performance in their presence?”

Learning Methods for Sequential Decision Making with Imperfect Representations

Shivaram Kalyanakrishnan. Ph.D. dissertation, University of Texas at Austin, 2011.

Shivaram Kalyanakrishnan 15/19

Practice =⇒ Imperfect Representations
State State Policy Representation
Task
Aliasing Space (Number of features)

Backgammon (T1992) Absent Discrete Neural network (198)

Job-shop scheduling (ZD1995) Absent Discrete Neural network (20)
Tetris (BT1906) Absent Discrete Linear (22)
Elevator dispatching (CB1996) Present Continuous Neural network (46)
Acrobot control (S1996) Absent Continuous Tile coding (4)
Dynamic channel allocation (SB1997) Absent Discrete Linear (100’s)
Active guidance of finless rocket (GM2003) Present Continuous Neural network (14)
Fast quadrupedal locomotion (KS2004) Present Continuous Parameterized policy (12)
Robot sensing strategy (KF2004) Present Continuous Linear (36)
Helicopter control (NKJS2004) Present Continuous Neural network (10)
Dynamic bipedal locomotion (TZS2004) Present Continuous Feedback control policy (2)
Adaptive job routing/scheduling (WS2004) Present Discrete Tabular (4)
Robot soccer keepaway (SSK2005) Present Continuous Tile coding (13)
Robot obstacle negotiation (LSYSN2006) Present Continuous Linear (10)
Optimized trade execution (NFK2007) Present Discrete Tabular (2-5)
Blimp control (RPHB2007) Present Continuous Gaussian Process (2)
9 × 9 Go (SSM2007) Absent Discrete Linear (≈1.5 million)
Ms. Pac-Man (SL2007) Absent Discrete Rule list (10)
Autonomic resource allocation (TJDB2007) Present Continuous Neural network (2)
General game playing (FB2008) Absent Discrete Tabular (part of state space)
Soccer opponent “hassling” (GRT2009) Present Continuous Neural network (9)
Adaptive epilepsy treatment (GVAP2008) Present Continuous Extremely rand. trees (114)
Computer memory scheduling (IMMC2008) Absent Discrete Tile coding (6)
Motor skills (PS2008) Present Continuous Motor primitive coeff. (100’s)
Combustion Control (HNGK2009) Present Continuous Parameterized policy (2-3)

Shivaram Kalyanakrishnan 16/19

Practice =⇒ Imperfect Representations
State State Policy Representation
Task
Aliasing Space (Number of features)

Backgammon (T1992) Absent Discrete Neural network (198)

Job-shop scheduling (ZD1995) Absent Discrete Neural network (20)
Tetris (BT1906) Absent Discrete Linear (22)
Elevator dispatching (CB1996) Present Continuous Neural network (46)
Acrobot control (S1996) Absent Continuous Tile coding (4)
Dynamic channel allocation (SB1997) Absent Discrete Linear (100’s)
Active guidance of finless rocket (GM2003) Present Continuous Neural network (14)
Fast quadrupedal locomotion (KS2004) Present Continuous Parameterized policy (12)
Robot sensing strategy (KF2004) Present Continuous Linear (36)
Helicopter control (NKJS2004) Present Continuous Neural network (10)
Dynamic bipedal locomotion (TZS2004) Present Continuous Feedback control policy (2)
Adaptive job routing/scheduling (WS2004) Present Discrete Tabular (4)
Robot soccer keepaway (SSK2005) Present Continuous Tile coding (13)
Robot obstacle negotiation (LSYSN2006) Present Continuous Linear (10)
Optimized trade execution (NFK2007) Present Discrete Tabular (2-5)
Blimp control (RPHB2007) Present Continuous Gaussian Process (2)
9 × 9 Go (SSM2007) Absent Discrete Linear (≈1.5 million)
Ms. Pac-Man (SL2007) Absent Discrete Rule list (10)
Autonomic resource allocation (TJDB2007) Present Continuous Neural network (2)
General game playing (FB2008) Absent Discrete Tabular (part of state space)
Soccer opponent “hassling” (GRT2009) Present Continuous Neural network (9)
Adaptive epilepsy treatment (GVAP2008) Present Continuous Extremely rand. trees (114)
Computer memory scheduling (IMMC2008) Absent Discrete Tile coding (6)
Motor skills (PS2008) Present Continuous Motor primitive coeff. (100’s)
Combustion Control (HNGK2009) Present Continuous Parameterized policy (2-3)

Shivaram Kalyanakrishnan 16/19

Practice =⇒ Imperfect Representations
State State Policy Representation
Task
Aliasing Space (Number of features)

Backgammon (T1992) Absent Discrete Neural network (198)

Job-shop scheduling (ZD1995) Absent Discrete Neural network (20)
Tetris (BT1906) Absent Discrete Linear (22)
Elevator dispatching (CB1996) Present Continuous Neural network (46)
Acrobot control (S1996) Absent Continuous Tile coding (4)
Dynamic channel allocation (SB1997) Absent Discrete Linear (100’s)
Active guidance of finless rocket (GM2003) Present Continuous Neural network (14)
Fast quadrupedal locomotion (KS2004) Present Continuous Parameterized policy (12)
Robot sensing strategy (KF2004) Present Continuous Linear (36)
Helicopter control (NKJS2004) Present Continuous Neural network (10)
Dynamic bipedal locomotion (TZS2004) Present Continuous Feedback control policy (2)
Adaptive job routing/scheduling (WS2004) Present Discrete Tabular (4)
Robot soccer keepaway (SSK2005) Present Continuous Tile coding (13)
Robot obstacle negotiation (LSYSN2006) Present Continuous Linear (10)
Optimized trade execution (NFK2007) Present Discrete Tabular (2-5)
Blimp control (RPHB2007) Present Continuous Gaussian Process (2)
9 × 9 Go (SSM2007) Absent Discrete Linear (≈1.5 million)
Ms. Pac-Man (SL2007) Absent Discrete Rule list (10)
Autonomic resource allocation (TJDB2007) Present Continuous Neural network (2)
General game playing (FB2008) Absent Discrete Tabular (part of state space)
Soccer opponent “hassling” (GRT2009) Present Continuous Neural network (9)
Adaptive epilepsy treatment (GVAP2008) Present Continuous Extremely rand. trees (114)
Computer memory scheduling (IMMC2008) Absent Discrete Tile coding (6)
Motor skills (PS2008) Present Continuous Motor primitive coeff. (100’s)
Combustion Control (HNGK2009) Present Continuous Parameterized policy (2-3)

Shivaram Kalyanakrishnan 16/19

Practice =⇒ Imperfect Representations
State State Policy Representation
Task
Aliasing Space (Number of features)

Backgammon (T1992) Absent Discrete Neural network (198)

Job-shop scheduling (ZD1995) Absent Discrete Neural network (20)
Tetris (BT1906) Absent Discrete Linear (22)
Elevator dispatching (CB1996) Present Continuous Neural network (46)
Acrobot control (S1996) Absent Continuous Tile coding (4)
Dynamic channel allocation (SB1997) Absent Discrete Linear (100’s)
Active guidance of finless rocket (GM2003) Present Continuous Neural network (14)
Fast quadrupedal locomotion (KS2004) Present Continuous Parameterized policy (12)
Robot sensing strategy (KF2004) Present Continuous Linear (36)
Helicopter control (NKJS2004) Present Continuous Neural network (10)
Dynamic bipedal locomotion (TZS2004) Present Continuous Feedback control policy (2)
Adaptive job routing/scheduling (WS2004) Present Discrete Tabular (4)
Robot soccer keepaway (SSK2005) Present Continuous Tile coding (13)
Robot obstacle negotiation (LSYSN2006) Present Continuous Linear (10)
Optimized trade execution (NFK2007) Present Discrete Tabular (2-5)
Blimp control (RPHB2007) Present Continuous Gaussian Process (2)
9 × 9 Go (SSM2007) Absent Discrete Linear (≈1.5 million)
Ms. Pac-Man (SL2007) Absent Discrete Rule list (10)
Autonomic resource allocation (TJDB2007) Present Continuous Neural network (2)
General game playing (FB2008) Absent Discrete Tabular (part of state space)
Soccer opponent “hassling” (GRT2009) Present Continuous Neural network (9)
Adaptive epilepsy treatment (GVAP2008) Present Continuous Extremely rand. trees (114)
Computer memory scheduling (IMMC2008) Absent Discrete Tile coding (6)
Motor skills (PS2008) Present Continuous Motor primitive coeff. (100’s)
Combustion Control (HNGK2009) Present Continuous Parameterized policy (2-3)

Perfect representations (fully observable, enumerable states) are impractical.

Shivaram Kalyanakrishnan 16/19

Today’s Class

1. Markov decision problems

2. Bellman’s (Optimality) Equations, planning and learning
3. Challenges
4. Summary

Shivaram Kalyanakrishnan 17/19

Summary

 Learning by trial and error to perform sequential decision making.

Shivaram Kalyanakrishnan 18/19

Summary

 Learning by trial and error to perform sequential decision making.

 Given an MDP (S, A, T , R, γ), we have to find a policy π : S → A that
yields high expected long-term reward from states.

Shivaram Kalyanakrishnan 18/19

Summary

 Learning by trial and error to perform sequential decision making.

 Given an MDP (S, A, T , R, γ), we have to find a policy π : S → A that
yields high expected long-term reward from states.
 An optimal value function V ∗ exists, and it induces an optimal policy π ∗
(several optimal policies might exist).

Shivaram Kalyanakrishnan 18/19

Summary

 Learning by trial and error to perform sequential decision making.

 Given an MDP (S, A, T , R, γ), we have to find a policy π : S → A that
yields high expected long-term reward from states.
 An optimal value function V ∗ exists, and it induces an optimal policy π ∗
(several optimal policies might exist).
 Under planning, we are given S, A, T , R, and γ. We may compute V ∗
and π ∗ using a dynamic programming algorithm such as policy iteration.

Shivaram Kalyanakrishnan 18/19

Summary

 Learning by trial and error to perform sequential decision making.

 Given an MDP (S, A, T , R, γ), we have to find a policy π : S → A that
yields high expected long-term reward from states.
 An optimal value function V ∗ exists, and it induces an optimal policy π ∗
(several optimal policies might exist).
 Under planning, we are given S, A, T , R, and γ. We may compute V ∗
and π ∗ using a dynamic programming algorithm such as policy iteration.
 In the learning context, we are given S, A, and γ: we may sample T and
R in a sequential manner. We can still converge to V ∗ and π ∗ by
applying a temporal difference learning method such as Q-learning.

Shivaram Kalyanakrishnan 18/19

Summary

 Learning by trial and error to perform sequential decision making.

 Given an MDP (S, A, T , R, γ), we have to find a policy π : S → A that
yields high expected long-term reward from states.
 An optimal value function V ∗ exists, and it induces an optimal policy π ∗
(several optimal policies might exist).
 Under planning, we are given S, A, T , R, and γ. We may compute V ∗
and π ∗ using a dynamic programming algorithm such as policy iteration.
 In the learning context, we are given S, A, and γ: we may sample T and
R in a sequential manner. We can still converge to V ∗ and π ∗ by
applying a temporal difference learning method such as Q-learning.
 Theory 6= Practice! In particular, convergence and optimality are difficult
to achieve when state spaces are large, and when state aliasing exists.

Shivaram Kalyanakrishnan 18/19

Summary

 Learning by trial and error to perform sequential decision making.

 Given an MDP (S, A, T , R, γ), we have to find a policy π : S → A that
yields high expected long-term reward from states.
 An optimal value function V ∗ exists, and it induces an optimal policy π ∗
(several optimal policies might exist).
 Under planning, we are given S, A, T , R, and γ. We may compute V ∗
and π ∗ using a dynamic programming algorithm such as policy iteration.
 In the learning context, we are given S, A, and γ: we may sample T and
R in a sequential manner. We can still converge to V ∗ and π ∗ by
applying a temporal difference learning method such as Q-learning.
 Theory 6= Practice! In particular, convergence and optimality are difficult
to achieve when state spaces are large, and when state aliasing exists.

Thank you!

Shivaram Kalyanakrishnan 18/19

References

Christopher J. C. H. Watkins and Peter Dayan, 1992. Q-Learning. Machine Learning,

8(3–4):279–292, 1992.

Leslie Pack Kaelbling, Michael L. Littman, and Andrew W. Moore, 1996. Reinforcement
Learning: A Survey. Journal of Artificial Intelligence Research, 4:237–285, 1996.

Jette Randløv and Preben Alstrøm, 1998. Learning to Drive a Bicycle using Reinforcement
Learning and Shaping. In Proceedings of the Fifteenth International Conference on Machine
Learning (ICML 1998), pp. 463–471, Morgan Kaufmann, 1998.

Richard S. Sutton and Andrew G. Barto, 1998. Reinforcement Learning: An Introduction.

MIT Press, 1998.

Pieter Abbeel, Adam Coates, Morgan Quigley, and Andrew Y. Ng, 2006. An Application of
Reinforcement Learning to Aerobatic Helicopter Flight. In Advances in Neural Information
Processing Systems 19, pp. 1–8, MIT Press, 2006.

Todd Hester, Michael Quinlan, and Peter Stone, 2010. Generalized Model Learning for
Reinforcement Learning on a Humanoid Robot. In Proceedings of the IEEE International
Conference on Robotics and Automation (ICRA 2010), pp. 2369–2374, IEEE, 2010.

Csaba Szepesvári, 2010. Algorithms for Reinforcement Learning. Morgan & Claypool, 2010.

Shivaram Kalyanakrishnan, 2011. Learning Methods for Sequential Decision Making with
Imperfect Representations. Ph.D. dissertation, published as UT Austin Computer Science
Technical Report TR-11-41, 2011.

Shivaram Kalyanakrishnan 19/19