Lecture 6: Value Function Approximation

Lecture 6: Value Function Approximation

David Silver
Lecture 6: Value Function Approximation

Outline

1 Introduction

2 Incremental Methods

3 Batch Methods
Lecture 6: Value Function Approximation
Introduction

Outline

1 Introduction

2 Incremental Methods

3 Batch Methods
Lecture 6: Value Function Approximation
Introduction

Large-Scale Reinforcement Learning

Reinforcement learning can be used to solve large problems, e.g.

Backgammon: 1020 states
Computer Go: 10170 states
Helicopter: continuous state space
Lecture 6: Value Function Approximation
Introduction

Large-Scale Reinforcement Learning

Reinforcement learning can be used to solve large problems, e.g.

Backgammon: 1020 states
Computer Go: 10170 states
Helicopter: continuous state space
How can we scale up the model-free methods for prediction and
control from the last two lectures?
Lecture 6: Value Function Approximation
Introduction

Value Function Approximation

So far we have represented value function by a lookup table

Every state s has an entry V (s)
Or every state-action pair s, a has an entry Q(s, a)
Problem with large MDPs:
There are too many states and/or actions to store in memory
It is too slow to learn the value of each state individually
Solution for large MDPs:
Estimate value function with function approximation

v̂ (s, w) ⇡ v⇡ (s)
or q̂(s, a, w) ⇡ q⇡ (s, a)

Generalise from seen states to unseen states

Update parameter w using MC or TD learning
Lecture 6: Value Function Approximation
Introduction

Types of Value Function Approximation

^
v(s,w) ^
q(s,a,w) ^
q(s,a … q(s,a
^
1,w) m,w)

w w w

s s a s
Lecture 6: Value Function Approximation
Introduction

Which Function Approximator?

There are many function approximators, e.g.

Linear combinations of features
Neural network
Decision tree
Nearest neighbour
Fourier / wavelet bases
...
Lecture 6: Value Function Approximation
Introduction

Which Function Approximator?

We consider di↵erentiable function approximators, e.g.

Linear combinations of features
Neural network
Decision tree
Nearest neighbour
Fourier / wavelet bases
...
Furthermore, we require a training method that is suitable for
non-stationary, non-iid data
Lecture 6: Value Function Approximation
Incremental Methods

Outline

1 Introduction

2 Incremental Methods

3 Batch Methods
Lecture 6: Value Function Approximation
Incremental Methods
Gradient Descent

Gradient Descent
!"#$%"&'(%")*+,#'-'!"#$%%" (%")*+,#'.+/0+,#
Let J(w) be a di↵erentiable function of
parameter vector w
Define!"#$%&'('%$&#()&*+$,*$#&&-&$$$$."'%"$
the gradient of J(w) to be
'*$%-/0'*,('-*$.'("$("#$1)*%('-*$
0 1
@J(w)
,22&-3'/,(-& %,*$0#$)+#4$(-$
B @w. 1 C
rw J(w) = @ .. C
B
%&#,(#$,*$#&&-&$1)*%('-*$$$$$$$$$$$$$
A
@J(w)
@wn
!"#$2,&(',5$4'11#&#*(',5$-1$("'+$#&&-&$
To find a local minimum of J(w)
1)*%('-*$$$$$$$$$$$$$$$$6$("#$7&,4'#*($
Adjust w in direction of -ve gradient
%,*$*-.$0#$)+#4$(-$)24,(#$("#$
'*(#&*,5$8,&',05#+$'*$("#$1)*%('-*$
1
w= ↵rw J(w)
,22&-3'/,(-&29,*4$%&'('%:;$$$$$$
where ↵ is a step-size parameter
Lecture 6: Value Function Approximation
Incremental Methods
Gradient Descent

Value Function Approx. By Stochastic Gradient Descent

Goal: find parameter vector w minimising mean-squared error

between approximate value fn v̂ (s, w) and true value fn v⇡ (s)
⇥ ⇤
J(w) = E⇡ (v⇡ (S) v̂ (S, w))2

Gradient descent finds a local minimum

1
w= ↵rw J(w)
2
= ↵E⇡ [(v⇡ (S) v̂ (S, w))rw v̂ (S, w)]

Stochastic gradient descent samples the gradient

w = ↵(v⇡ (S) v̂ (S, w))rw v̂ (S, w)

Expected update is equal to full gradient update

Lecture 6: Value Function Approximation
Incremental Methods
Linear Function Approximation

Feature Vectors

Represent state by a feature vector

0 1
x1 (S)
B C
x(S) = @ ... A
xn (S)

For example:
Distance of robot from landmarks
Trends in the stock market
Piece and pawn configurations in chess
Lecture 6: Value Function Approximation
Incremental Methods
Linear Function Approximation

Linear Value Function Approximation

Represent value function by a linear combination of features
n
X
v̂ (S, w) = x(S)> w = xj (S)wj
j=1

Objective function is quadratic in parameters w

h i
J(w) = E⇡ (v⇡ (S) x(S)> w)2

Stochastic gradient descent converges on global optimum

Update rule is particularly simple
rw v̂ (S, w) = x(S)
w = ↵(v⇡ (S) v̂ (S, w))x(S)
Update = step-size ⇥ prediction error ⇥ feature value
Lecture 6: Value Function Approximation
Incremental Methods
Linear Function Approximation

Table Lookup Features

Table lookup is a special case of linear value function

approximation
Using table lookup features
0 1
1(S = s1 )
B .. C
xtable (S) = @ . A
1(S = sn )

Parameter vector w gives value of each individual state

0 1 0 1
1(S = s1 ) w1
B .
.. C B . C
v̂ (S, w) = @ A · @ .. A
1(S = sn ) wn
Lecture 6: Value Function Approximation
Incremental Methods
Incremental Prediction Algorithms

Incremental Prediction Algorithms

Have assumed true value function v⇡ (s) given by supervisor

But in RL there is no supervisor, only rewards
In practice, we substitute a target for v⇡ (s)
For MC, the target is the return Gt

w = ↵(Gt v̂ (St , w))rw v̂ (St , w)

For TD(0), the target is the TD target Rt+1 + v̂ (St+1 , w)

w = ↵(Rt+1 + v̂ (St+1 , w) v̂ (St , w))rw v̂ (St , w)

For TD( ), the target is the -return Gt

w = ↵(Gt v̂ (St , w))rw v̂ (St , w)

Lecture 6: Value Function Approximation
Incremental Methods
Incremental Prediction Algorithms

Monte-Carlo with Value Function Approximation

Return Gt is an unbiased, noisy sample of true value v⇡ (St )

Can therefore apply supervised learning to “training data”:

hS1 , G1 i, hS2 , G2 i, ..., hST , GT i

For example, using linear Monte-Carlo policy evaluation

w = ↵(Gt v̂ (St , w))rw v̂ (St , w)

= ↵(Gt v̂ (St , w))x(St )

Monte-Carlo evaluation converges to a local optimum

Even when using non-linear value function approximation
Lecture 6: Value Function Approximation
Incremental Methods
Incremental Prediction Algorithms

TD Learning with Value Function Approximation

The TD-target Rt+1 + v̂ (St+1 , w) is a biased sample of true

value v⇡ (St )
Can still apply supervised learning to “training data”:

hS1 , R2 + v̂ (S2 , w)i, hS2 , R3 + v̂ (S3 , w)i, ..., hST 1 , RT i

For example, using linear TD(0)

w = ↵(R + v̂ (S 0 , w) v̂ (S, w))rw v̂ (S, w)

= ↵ x(S)

Linear TD(0) converges (close) to global optimum

Lecture 6: Value Function Approximation
Incremental Methods
Incremental Prediction Algorithms

TD( ) with Value Function Approximation

The -return Gt is also a biased sample of true value v⇡ (s)
Can again apply supervised learning to “training data”:
D E D E D E
S1 , G1 , S2 , G2 , ..., ST 1 , GT 1

Forward view linear TD( )

w = ↵(Gt v̂ (St , w))rw v̂ (St , w)
= ↵(Gt v̂ (St , w))x(St )
Backward view linear TD( )
t = Rt+1 + v̂ (St+1 , w) v̂ (St , w)
Et = Et 1 + x(St )
w = ↵ t Et
Lecture 6: Value Function Approximation
Incremental Methods
Incremental Prediction Algorithms

TD( ) with Value Function Approximation

The -return Gt is also a biased sample of true value v⇡ (s)
Can again apply supervised learning to “training data”:
D E D E D E
S1 , G1 , S2 , G2 , ..., ST 1 , GT 1

Forward view linear TD( )

w = ↵(Gt v̂ (St , w))rw v̂ (St , w)
= ↵(Gt v̂ (St , w))x(St )
Backward view linear TD( )
t = Rt+1 + v̂ (St+1 , w) v̂ (St , w)
Et = Et 1 + x(St )
w = ↵ t Et
Forward view and backward view linear TD( ) are equivalent
Lecture 6: Value Function Approximation
Incremental Methods
Incremental Control Algorithms

Control with Value Function Approximation

q
w =q
π

Starting w qw ≈ q*

)
dy(q w
ε- gree
π =

Policy evaluation Approximate policy evaluation, q̂(·, ·, w) ⇡ q⇡

Policy improvement ✏-greedy policy improvement
Lecture 6: Value Function Approximation
Incremental Methods
Incremental Control Algorithms

Action-Value Function Approximation

Approximate the action-value function

q̂(S, A, w) ⇡ q⇡ (S, A)

Minimise mean-squared error between approximate

action-value fn q̂(S, A, w) and true action-value fn q⇡ (S, A)
⇥ ⇤
J(w) = E⇡ (q⇡ (S, A) q̂(S, A, w))2

Use stochastic gradient descent to find a local minimum

1
rw J(w) = (q⇡ (S, A) q̂(S, A, w))rw q̂(S, A, w)
2
w = ↵(q⇡ (S, A) q̂(S, A, w))rw q̂(S, A, w)
Lecture 6: Value Function Approximation
Incremental Methods
Incremental Control Algorithms

Linear Action-Value Function Approximation

Represent state and action by a feature vector
0 1
x1 (S, A)
B .. C
x(S, A) = @ . A
xn (S, A)
Represent action-value fn by linear combination of features
n
X
q̂(S, A, w) = x(S, A)> w = xj (S, A)wj
j=1

Stochastic gradient descent update

rw q̂(S, A, w) = x(S, A)
w = ↵(q⇡ (S, A) q̂(S, A, w))x(S, A)
Lecture 6: Value Function Approximation
Incremental Methods
Incremental Control Algorithms

Incremental Control Algorithms

Like prediction, we must substitute a target for q⇡ (S, A)
For MC, the target is the return Gt
w = ↵(Gt q̂(St , At , w))rw q̂(St , At , w)
For TD(0), the target is the TD target Rt+1 + Q(St+1 , At+1 )
w = ↵(Rt+1 + q̂(St+1 , At+1 , w) q̂(St , At , w))rw q̂(St , At , w)
For forward-view TD( ), target is the action-value -return
w = ↵(qt q̂(St , At , w))rw q̂(St , At , w)
For backward-view TD( ), equivalent update is

t = Rt+1 + q̂(St+1 , At+1 , w) q̂(St , At , w)

Et = Et 1 + rw q̂(St , At , w)
w = ↵ t Et
Lecture 6: Value Function Approximation
Incremental Methods
Mountain Car

Linear Sarsa with Coarse Coding in Mountain Car

Lecture 6: Value Function Approximation
Incremental Methods
Mountain Car

Linear Sarsa with Radial Basis Functions in Mountain Car

Lecture 6: Value Function Approximation
Incremental Methods
Mountain Car

Study of : Should We Bootstrap?

Lecture 6: Value Function Approximation
Incremental Methods
Convergence

Baird’s Counterexample
Lecture 6: Value Function Approximation
Incremental Methods
Convergence

Parameter Divergence in Baird’s Counterexample

Lecture 6: Value Function Approximation
Incremental Methods
Convergence

Convergence of Prediction Algorithms

On/O↵-Policy Algorithm Table Lookup Linear Non-Linear

MC 3 3 3
On-Policy
TD(0) 3 3 7
TD( ) 3 3 7
MC 3 3 3
O↵-Policy
TD(0) 3 7 7
TD( ) 3 7 7
Lecture 6: Value Function Approximation
Incremental Methods
Convergence

Gradient Temporal-Di↵erence Learning

TD does not follow the gradient of any objective function

This is why TD can diverge when o↵-policy or using
non-linear function approximation
Gradient TD follows true gradient of projected Bellman error

On/O↵-Policy Algorithm Table Lookup Linear Non-Linear

MC 3 3 3
On-Policy
TD 3 3 7
Gradient TD 3 3 3
MC 3 3 3
O↵-Policy
TD 3 7 7
Gradient TD 3 3 3
Lecture 6: Value Function Approximation
Incremental Methods
Convergence

Convergence of Control Algorithms

Algorithm Table Lookup Linear Non-Linear

Monte-Carlo Control 3 (3) 7
Sarsa 3 (3) 7
Q-learning 3 7 7
Gradient Q-learning 3 3 7

(3) = chatters around near-optimal value function

Lecture 6: Value Function Approximation
Batch Methods

Outline

1 Introduction

2 Incremental Methods

3 Batch Methods
Lecture 6: Value Function Approximation
Batch Methods

Batch Reinforcement Learning

Gradient descent is simple and appealing

But it is not sample efficient
Batch methods seek to find the best fitting value function
Given the agent’s experience (“training data”)
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

Least Squares Prediction

Given value function approximation v̂ (s, w) ⇡ v⇡ (s)

And experience D consisting of hstate, valuei pairs

D = {hs1 , v1⇡ i, hs2 , v2⇡ i, ..., hsT , vT⇡ i}

Which parameters w give the best fitting value fn v̂ (s, w)?

Least squares algorithms find parameter vector w minimising
sum-squared error between v̂ (st , w) and target values vt⇡ ,
T
X
LS(w) = (vt⇡ v̂ (st , w))2
t=1
⇥ ⇤
= ED (v ⇡ v̂ (s, w))2
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

Stochastic Gradient Descent with Experience Replay

Given experience consisting of hstate, valuei pairs

D = {hs1 , v1⇡ i, hs2 , v2⇡ i, ..., hsT , vT⇡ i}

Repeat:
1 Sample state, value from experience

hs, v ⇡ i ⇠ D

2 Apply stochastic gradient descent update

w = ↵(v ⇡ v̂ (s, w))rw v̂ (s, w)

Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

Stochastic Gradient Descent with Experience Replay

Given experience consisting of hstate, valuei pairs

D = {hs1 , v1⇡ i, hs2 , v2⇡ i, ..., hsT , vT⇡ i}

Repeat:
1 Sample state, value from experience

hs, v ⇡ i ⇠ D

2 Apply stochastic gradient descent update

w = ↵(v ⇡ v̂ (s, w))rw v̂ (s, w)

Converges to least squares solution

w⇡ = argmin LS(w)
w
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

Experience Replay in Deep Q-Networks (DQN)

DQN uses experience replay and fixed Q-targets

Take action at according to ✏-greedy policy
Store transition (st , at , rt+1 , st+1 ) in replay memory D
Sample random mini-batch of transitions (s, a, r , s 0 ) from D
Compute Q-learning targets w.r.t. old, fixed parameters w
Optimise MSE between Q-network and Q-learning targets
"✓ ◆2 #
Li (wi ) = Es,a,r ,s 0 ⇠Di r+ max
0
Q(s 0 , a0 ; wi ) Q(s, a; wi )
a

Using variant of stochastic gradient descent

Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

DQN in Atari

End-to-end learning of values Q(s, a) from pixels s

Input state s is stack of raw pixels from last 4 frames
Output is Q(s, a) for 18 joystick/button positions
Reward is change in score for that step

Network architecture and hyperparameters fixed across all games

Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

DQN Results in Atari

Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

How much does DQN help?

Replay Replay No replay No replay

Fixed-Q Q-learning Fixed-Q Q-learning
Breakout 316.81 240.73 10.16 3.17
Enduro 1006.3 831.25 141.89 29.1
River Raid 7446.62 4102.81 2867.66 1453.02
Seaquest 2894.4 822.55 1003 275.81
Space Invaders 1088.94 826.33 373.22 301.99
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

Linear Least Squares Prediction

Experience replay finds least squares solution

But it may take many iterations
Using linear value function approximation v̂ (s, w) = x(s)> w
We can solve the least squares solution directly
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

Linear Least Squares Prediction (2)

At minimum of LS(w), the expected update must be zero

ED [ w] = 0
T
X
↵ x(st )(vt⇡ x(st )> w) = 0
t=1
T
X T
X
x(st )vt⇡ = x(st )x(st )> w
t=1 t=1
T
! 1 T
X X
>
w= x(st )x(st ) x(st )vt⇡
t=1 t=1

For N features, direct solution time is O(N 3 )

Incremental solution time is O(N 2 ) using Shermann-Morrison
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

Linear Least Squares Prediction Algorithms

We do not know true values vt⇡

In practice, our “training data” must use noisy or biased
samples of vt⇡
LSMC Least Squares Monte-Carlo uses return
vt⇡ ⇡ Gt
LSTD Least Squares Temporal-Di↵erence uses TD target
vt⇡ ⇡ Rt+1 + v̂ (St+1 , w)
LSTD( ) Least Squares TD( ) uses -return
vt⇡ ⇡ Gt

In each case solve directly for fixed point of MC / TD / TD( )

Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

Linear Least Squares Prediction Algorithms (2)

T
X
LSMC 0= ↵(Gt v̂ (St , w))x(St )
t=1
T
! 1 T
X X
>
w= x(St )x(St ) x(St )Gt
t=1 t=1
T
X
LSTD 0= ↵(Rt+1 + v̂ (St+1 , w) v̂ (St , w))x(St )
t=1
T
! 1 T
X X
>
w= x(St )(x(St ) x(St+1 )) x(St )Rt+1
t=1 t=1
T
X
LSTD( ) 0= ↵ t Et
t=1
T
! 1 T
X X
>
w= Et (x(St ) x(St+1 )) Et Rt+1
t=1 t=1
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Prediction

Convergence of Linear Least Squares Prediction Algorithms

On/O↵-Policy Algorithm Table Lookup Linear Non-Linear

MC 3 3 3
LSMC 3 3 -
On-Policy
TD 3 3 7
LSTD 3 3 -
MC 3 3 3
O↵-Policy
LSMC 3 3 -
TD 3 7 7
LSTD 3 3 -
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Control

Least Squares Policy Iteration

q
w =q
π

Starting w qw ≈ q*

)
(q w
g r eedy
π =

Policy evaluation Policy evaluation by least squares Q-learning

Policy improvement Greedy policy improvement
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Control

Least Squares Action-Value Function Approximation

Approximate action-value function q⇡ (s, a)

using linear combination of features x(s, a)

q̂(s, a, w) = x(s, a)> w ⇡ q⇡ (s, a)

Minimise least squares error between q̂(s, a, w) and q⇡ (s, a)

from experience generated using policy ⇡
consisting of h(state, action), valuei pairs

D = {h(s1 , a1 ), v1⇡ i, h(s2 , a2 ), v2⇡ i, ..., h(sT , aT ), vT⇡ i}

Lecture 6: Value Function Approximation
Batch Methods
Least Squares Control

Least Squares Control

For policy evaluation, we want to efficiently use all experience

For control, we also want to improve the policy
This experience is generated from many policies
So to evaluate q⇡ (S, A) we must learn o↵-policy
We use the same idea as Q-learning:
Use experience generated by old policy
St , At , Rt+1 , St+1 ⇠ ⇡old
Consider alternative successor action A0 = ⇡new (St+1 )
Update q̂(St , At , w) towards value of alternative action
Rt+1 + q̂(St+1 , A0 , w))
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Control

Least Squares Q-Learning

Consider the following linear Q-learning update

= Rt+1 + q̂(St+1 , ⇡(St+1 ), w) q̂(St , At , w)

w = ↵ x(St , At )

LSTDQ algorithm: solve for total update = zero

T
X
0= ↵(Rt+1 + q̂(St+1 , ⇡(St+1 ), w) q̂(St , At , w))x(St , At )
t=1
T
! 1 T
X X
>
w= x(St , At )(x(St , At ) x(St+1 , ⇡(St+1 ))) x(St , At )Rt+1
t=1 t=1
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Control

Least Squares Policy Iteration Algorithm

The following pseudocode uses LSTDQ for policy evaluation

It repeatedly re-evaluates experience D with di↵erent policies
function LSPI-TD(D, ⇡0 )
⇡0 ⇡0
repeat
⇡ ⇡0
Q LSTDQ(⇡, D)
for all s 2 S do
⇡ 0 (s) argmax Q(s, a)
a2A
end for
until (⇡ ⇡ ⇡ 0 )
return ⇡
end function
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Control

Convergence of Control Algorithms

Algorithm Table Lookup Linear Non-Linear

Monte-Carlo Control 3 (3) 7
Sarsa 3 (3) 7
Q-learning 3 7 7
LSPI 3 (3) -

(3) = chatters around near-optimal value function

Lecture 6: Value Function Approximation
Batch Methods
Least Squares Control

Least-Squares Policy Iteration

Chain Walk Example
0.1 0.1
0.1

0.1
0.9 0.9 0.9

R R R R
L
1 2 3 4 0.9
0.9 L L L

0.9 0.9 0.9

0.1
0.1
0.1 0.1

r=0 r=1 r=1 r=0

Figure 9: The problematic MDP.

Consider the 50 state version of this problem
where s is the state number. LSPI was applied on the same problem using the same basis
Reward
functions +1forineachstates
repeated 10actions
of the two and 41, 0 each
so that elsewhere
action gets its own parameters:10
Optimal policy: R (1-9),L I(a (10-25),
= L) × 1 R
(26-41), L (42, 50)
 I(a = L) × s 
Features: 10 evenly spaced  Gaussians 
 I(a = L) × s2 ( = 4) for each action
φ(s, a) =  
 I(a = R) × 1  .
Experience: 10,000 steps from random
I(a = R) × s 
 walk policy

I(a = R) × s2
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Control

Least-Squares Policy Iteration

LSPI in Chain Walk: Action-Value Function
1.5 1.5
1
1 0.5
0.5 0
−0.5
0
5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Iteration1 Iteration2

4 4

2
2

0
0
5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Iteration3 Iteration4

4 4

2 2

0 0
5 10 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Iteration5 Iteration6

0
5 10 15 20 25 30 35 40 45 50
Iteration7
 2 2
Lecture 6: Value Function Approximation
Batch Methods
0 0
5 10Control
Least Squares 15 20 25 30 35 40 45 50 5 10 15 20 25 30 35 40 45 50
Iteration5 Iteration6

4
LSPI in Chain Walk: Policy
2

0
5 10 15 20 25 30 35 40 45 50
Iteration7

10 20 30 40 50 10 20 30 40 50
Iteration1 Iteration2

10 20 30 40 50 10 20 30 40 50
Iteration3 Iteration4

10 20 30 40 50 10 20 30 40 50
Iteration5 Iteration6

10 20 30 40 50
Iteration7

Figure 13: LSPI iterations on a 50-state chain with a radial basis function approximator
(reward only in states 10 and 41). Top: The state-action value function of the
policy being evaluated in each iteration (LSPI approximation - solid lines; exact
Lecture 6: Value Function Approximation
Batch Methods
Least Squares Control

Questions?