1

CE7454 : Deep Learning for Data Science

Lecture 15: Deep Reinforcement Learning

Semester 1 2019/20

Xavier Bresson
School of Computer Science and Engineering
Data Science and AI Research Centre
Nanyang Technological University (NTU), Singapore

Xavier Bresson 1
 2

Material :
David Silver :
Tutorial on Deep Reinforcement Learning, ICML’16
UCL Course on Reinforcement Learning, 2015
Pieter Abbeel :
Talk at Simons Institute for the Theory of Computing, Berkeley, 2017
Mario Martin :
Slides on Policy Search: Actor-Critic and Gradient Policy search, 2019

Xavier Bresson 2
 3

Outline

RL and Deep Learning

Agent, Environment and MDP
Policy, Value Function and Model
Optimal Value Function and Policy
Deep Q-Learning (DQN)
Policy Networks
Actor-Critic Algorithm
RL and Supervised Learning
Learning and Planning

Xavier Bresson 3
 4

Outline

RL and Deep Learning

Agent, Environment and MDP
Policy, Value Function and Model
Optimal Value Function and Policy
Deep Q-Learning (DQN)
Policy Networks
Actor-Critic Algorithm
RL and Supervised Learning
Learning and Planning

Xavier Bresson 4
 5

Deep Learning

DL defines a framework to learn the best possible representation of data that can solve tasks
like classification, regression, recommender systems, etc.
DL is a supervised learning technique :
Data are all labeled.
Research and industrial breakthrough applications in CV (autonomous cars, etc), Speech
Recognition (speech-to-text and inversely), NLP (machine translation, etc).

Xavier Bresson 5
 6

Deep Learning

Potential revolutionary applications in other fields like physics (simulation acceleration, particle
detection with complex geometry and 108 sensors, generative models for machine calibration)
and chemistry (drugs and materials design).
Learn a deep compositional parametric functions by backpropagation on the loss function.
Loss functions are continuous and differentiable.

Quark and gluon field

Particle jet Molecule design
fluctuations

Xavier Bresson 6
 7

Reinforcement Learning

RL defines a framework to learn the best possible sequence of decisions in a given environment
that can solve tasks like playing Go/Chess, making autonomous robots, recommending ads on
internet, etc.

RL is the science of decision-making (D. Silver) :

Goal is to maximize future reward.
Decisions may have long-term consequences :
AlphaGo vs Lee Sedol with move 37 in game 2.
High positive and negative rewards are not instantaneous.
Reward may be delayed for higher values (giving up short-term rewards).

RL is a causal framework.
Time is important – RL is a sequential decision-making technique.

Xavier Bresson 7
 8

Reinforcement Learning

RL is different from supervised learning, but has also similarities (later discussed).

Unlike DL, no real-world breakthrough applications of RL (except super-human game players).

Applications to real-world problems are limited, but it might change quickly (DeepMind).

Applications :
Playing games (super-human performances): Go, Chess, Poker, Atari, StarCraft II
Exploring virtual worlds : Maze
Robots (still preliminary) : Simulation and real autonomous robots
Learning to learn (meta-learning): Compute simultaneously architectures and weights.
End of humans designing deep learning architectures ?

Xavier Bresson 8
 9

Reinforcement Learning

Applications :

Helicopters that learn to be autonomous Robots that learn to grasp objects

(Stanford) (Google)

Xavier Bresson 9
 10

Reinforcement Learning

Applications :

Historical AI milestone AlphaStar for StarCraft II – Jan 2019

Game of Go – March 2016 (DeepMind)
(DeepMind)
AlphaStar beat 5-0 Team Liquid’s Grzegorz
"MaNa" Komincz, one of the world’s strongest
professional StarCraft players
https://deepmind.com/blog/article/alphastar-
mastering-real-time-strategy-game-starcraft-ii

Xavier Bresson 10
 11

Brain Inspiration

The human brain may use the cerebellum for

supervised learning representation of the
world data.

The human brain may use the cortex for

unsupervised learning representation of the
world data.

The human brain may use the basal ganglia

for making decisions on problems. John Sowa

Xavier Bresson 11
 12

Deep RL
Reinforcement learning has developed powerful algorithms to solve control problems :
Dynamic programming 1953 (Richard Bellman) :
Solve a difficult problem by breaking it into simpler sub-problems in a recursive
manner.
DP can solve optimally low-dimensional control problems by looking at the whole
search space.
DP becomes intractable in high-dimensional search space.
Solution :
Combine DL + RL = Deep RL

Best Best decision Best decision

representation maker in low- maker in high-
learning dim space. dim space.

Each move in Go leads to ~250 possible moves.

The exploration space has 10170 paths.

Xavier Bresson 12
 13

Outline

RL and Deep Learning

Agent, Environment and MDP
Policy, Value Function and Model
Optimal Value Function and Policy
Deep Q-Learning (DQN)
Policy Networks
Actor-Critic Algorithm
RL and Supervised Learning
Learning and Planning

Xavier Bresson 13
 14

Agent and Environment

RL framework :
Splitting agent and environment :
The environment/world can be anything, from simple world like Go board to
complex world like Earth.
The agent can make actions/decisions.
Each action influences/changes the state of the environment (where the
agent lives).
Each action produces a scalar reward from the environment.

Goal of RL :
Design an agent that acts optimally to optimize reward.

Xavier Bresson 14
 15

Agent and Environment

RL framework :
At each time step t :
The agent:
Observation ot Action at
Receives an observation ot and
executes an action at.
Observes ot+1 the result of its action. Agent
Receives a scalar reward rt.
The environment :
Receives an action at.
Produces two outputs: Observation ot+1
Reward rt
Observation ot+1 Environment
Reward rt

Xavier Bresson 15
 16

Episode, State and Environment

An episode/trajectory/experience et is a sequence of observations, actions, rewards :

et = o1, a1, o2, r1, a2, o3, r2, … , ot, at, ot+1, rt

A state st is a summary of the episode/experience : State st Action at

st = f(et)
= f(o1, a1, o2, r1, a2, o3, r2, … , ot, at, ot+1, rt)
Agent

If the environment is fully observable :

Go, Chess, but not Poker
st = f(et) = f(o1, a1, o2, r1, a2, o3, r2, … , ot, at, ot+1, rt) = ot
State st+1
⇒ st = ot Reward rt
An episode is written as: Environment

et = s1, a1, s2, r1, a2, s3, r2, … , st, at, st+1, rt
At time step t :
st, at, st+1, rt

Xavier Bresson 16
 17

MDP

A fully observable environment is defined by a Markov Decision Process (MDP) :

The current state st completely characterizes the future, independently of the past.

A state st is Markov if the state transition probability P is defined as :

P(st+1|st) = P(s+1|st, st-1,…,s1)
The current state st captures all relevant information from the history st-1,…,s1.
The history is not useful and can be discarded.

A Markov process/chain is a memoryless random process.

A sequence of states s1,s2,…,st with Markov property.

Xavier Bresson 17
 18

MDP

A Markov decision process (MDP) :

A Markov process with decisions/actions.

The state transition probability matrix is defined as :

P(st+1|st,at)
It encodes the dynamics/behavior of the environment where the agent lives.

Xavier Bresson 18
 19

Example
Pong can be cast as a MDP :
A MDP can be visualized as a graph where :
Each node is either a game state st or an action at.
Each edge is a probabilistic transition given either by the state transition probability
matrix P(st+1|st,at), or the policy function π(at|st) (later discussed).

π(a1|s1)
s1 a1
π(a2|s1) P(s2|s1,a1)

P(s1|s3,a3) a2 s2

P(s3|s1,a1) π(a4|s2)
a3 P(s3|s1,a2)

s3 a4
π(a3|s3)
P(s3|s2,a4)
PONG : A game state st is the
image of the game at time state t.
An action at is a move, MDP of PONG
either Up or Down.
Xavier Bresson 19
 20

Outline

RL and Deep Learning

Agent, Environment and MDP
Policy, Value Function and Model
Optimal Value Function and Policy
Deep Q-Learning (DQN)
Policy Networks
Actor-Critic Algorithm
RL and Supervised Learning
Learning and Planning

Xavier Bresson 20
 21

Properties of RL

The goal of RL is to act optimally to maximize a value function that represents the total
reward of a sequence of actions.

There are three types of RL problems, based on three fundamental properties of RL :

Policy function :
Agent’s behavior function (select the agent’s action).
Value function:
Function that quantifies how good or bad is each state or action in state.
Model function :
Agent’s representation of the environment (for planning/reasoning).

We will use a simple example to illustrate these properties:

Maze

Xavier Bresson 21
 22

Policy

A policy encodes the agent’s behavior :

It is a map from states to actions.
Deterministic policy :
a = π(s)
Given deterministic π, a new episode can be drawn as follows :
et = s1, a1 = π(s1), s2 ~ P(s|s1,a1), r1, a2 = π(s2), s3 ~ P(s|s2,a2), r2,
… st ~ P(s|st-1,at-1), at = π(st), st+1 ~ P(s|st,at), rt

r1(s1,a1,s2) r2(s2,a2,s3) rt(st,at,st+1)

π(a1) P(s2|s1,a1) π(a2) P(s3|s2,a2) π(at) P(st+1|st,at)

s1 a1 s2 a2 s3 … at st+1

Xavier Bresson 22
 23

Policy

Stochastic policy :
π(a|s) = Probability of action a in state s
Given stochastic π, a new episode can be drawn as follows :
et = s1, a1 ~ π(a|s1), s2 ~ P(s|s1,a1), r1, a2 ~ π(a|s2), s3 ~ P(s|s2,a2), r2,
… st ~ P(s|st-1,at-1), at ~ π(a|st), st+1 ~ P(s|st,at), rt
From a given state st, there are multiple possible actions, unlike for the
deterministic policy.

r1(s1,a1,s2) r2(s2,a2,s3) rt(st,at,st+1)

π(a1|s1) P(s2|s1,a1) π(a2|s2) P(s3|s2,a2) π(at|st) P(st+1|st,at)

s1 a1 s2 a2 s3 … at st+1

Xavier Bresson 23
 24

Policy

How to sample at ~ π(a|st) and st ~ P(s|st-1,at-1) ?

Bernoulli sampling :
A random draw with exactly K possible outcomes, in which the probability Pr of
having X=xk depends on a stationary distribution (i.e. the distribution stays
unchanged at each time step).
NumPy/PyTorch :
s = np.random.choice( p=Pr(X) )
s = torch.distributions.Categorical( p=Pr(X) ).sample()

Pr

X
x1 x2 x3 xK

Xavier Bresson 24
 25

Policy
Given probability functions π(a|s) and P(a|s),
A new episode et can be drawn by rolling out/sampling π and P :
et = s1, a1 ~ π(a|s1), s2 ~ P(s|s1,a1), r1, a2 ~ π(a|s2), s3 ~ P(s|s2,a2), r2,
st+1
… st ~ P(s|st-1,at-1), at ~ π(a|st), st+1 ~ P(s|st,at), rt
at
st+1
s3
at
a2 st+1
s3
s2 at
a2 st+1
s3
a1 at
s2 a2
st+1
s3
at
s1 a1 s2 a2 s3 … at
st+1

st+1
s2 a2 s3
π(a1|s1) a1 at
st+1

a2 s3
s2 at
P(s2|s1,a1) st+1
s3
π(a2|s2) a2 at
st+1
s3 at
P(s3|s2,a2) st+1
Xavier Bresson P(st+1|st,at) 25
 26

Policy

Maze example :
Actions are : Up, Down, Left, Right
States are : Agent’s locations

Arrows represent optimal policy π(s)

for each state s
Xavier Bresson 26
 27

Model

Maze example :
Actions are : Up, Down, Left, Right
States are : Agent’s locations
Model is :
Dynamics of the environment modeled by P(s|st,at)
Immediate reward rt(st,at,st+1) given by the environment

Dynamics : Grid layout represents the transition

model P
(same probability value for all state-action)

Reward : Number -1 represents immediate reward r

from each state s
(same reward value for all actions a)

Xavier Bresson 27
 28

State-Value Function

The state value function Q(s) is the expected future reward for being in state s.
It is used to evaluate the goodness/badness of each state s.
It acts as the label information in supervised learning (later discussed).

Let us compute the value function Q in state st considering only one episode :
et = st, at ~ π(a|st), st+1 ~ P(s|st,at), rt, at+1 ~ π(a|st+1), st+2 ~ P(s|st+1,at+1), rt+1, …
Q(st) = rt + γ.rt+1 + γ2.rt+2 + … = Σk=0∞ γk.rt+k

Return for Immediate Sum of discounted

being in state st reward future rewards

The value function Q(st) is the total discounted reward (a.k.a. return) we will receive
for being in state st.
rt(st,at,st+1) rt+1(st+2,at+1,st+1)

π(at|st) P(st+1|st,at) π(at+1|st+1) P(st+2|st+1,at+1)

st at st+1 at+1 st+2 …
Xavier Bresson 28
 29

State-Value Function
Discounted reward using hyper-parameter γ :
Q(st) = rt + γ.rt+1 + γ2.rt+2 + … = Σk=0∞ γk.rt+k
Exponential decay factor γk :
0<γ<1
γ=0.9 ⇒ γk ≈ 0 for k ≥ 50 : short-term consequences
γ=0.99 ⇒ γk ≈ 0 for k ≥ 500 : long-term consequences (harder to learn)
Justifications of γ :
Math :
Avoid infinite total reward with value γ<1.
Model uncertainty about the policy network :
The action taken is not absolutely certain to be optimal, so it is better to
discount its future reward.
Short-term return is usually preferred than long-term (e.g. Finance).

Xavier Bresson 29
 30

State-Value Function
Maze example :
Actions are : Up, Down, Left, Right
States are : Agent’s locations
Model is :
Dynamics of the environment modeled by P(s|st,at)
Immediate reward rt(st,at,st+1) = -1 given by the environment

Numbers represent optimal value

function Q(s) (w/ γ=1) for each state s.

Xavier Bresson 30
 31

State-Value Function

Let us compute the value function Q in state st considering all possible episodes drawn
from π and P :
et = st, at ~ π(a|st), st+1 ~ P(s|st,at), rt, at+1 ~ π(a|st+1), st+2 ~ P(s|st+1,at+1), rt+1, …
Q(st) = !π,P ( rt + γ.rt+1 + γ2.rt+2 + … )
Q(st) is the expected total discounted reward for being in state st.

rt(st,at,st+1) rt+1(st+2,at+1,st+1)

π(at|st) P(st+1|st,at) π(at+1|st+1) P(st+2|st+1,at+1)

st at st+1 at+1 st+2 …

Xavier Bresson 31
 32

State-Value Function

Let us write the Bellman equation of the value function Q :

Q(st) = !π,P ( rt + γ.rt+1 + γ2.rt+2 + … )
= !π,P ( rt + !π,P ( γ.rt+1 + γ2.rt+2 + … ) )
= !π,P ( rt + γ. !π,P ( rt+1 + γ.rt+2 + … ) )
= !π,P ( rt + γ. Q(st+1) )
Recursive formulation (dynamic programming)
The sum of total reward Q for being in state st can be decomposed into 2 parts :
The immediate reward rt we get after one time step.
The discounted sum of future rewards γ.Q(st+1) we will receive.
It is the expectation over all possible states st+1.

Xavier Bresson 32
 33

State-Value Function

The Bellman equation of the value function is stationary, i.e. independent of the time step t :
Let st, at, st+1, rt be re-written as
s, a, s’, r
Let Q(st) = !π,P ( rt + γ. Q(st+1) ) be re-written as
Q(s) = !π,P ( r(s,a,s’)+ γ. Q(s’) )
= Σa π(a|s) Σs’ P(s’|a,s) ( r(s,a,s’) + γ. Q(s’) )

Probability of action Transition probability r(s,a,s’)

a in state s from action a in state
s to state s’ s’
P(s’|s,a)
r(s,a,s’)
a

…
π(a|s)
s’
P(s’|s,a) r(s,a,s’)
s
P(s’|s,a) s’
π(a|s) r(s,a,s’)
a

P(s’|s,a) s’

Xavier Bresson 33
 34

Action-Value Function

The state-value function Q(st) can be extended to an action-value function Q(at|st) :

Q(at|st) = Future reward (good or bad) of selecting action at in state st.

Let us compute the value function Q of taking action at in state st considering only one episode :
et = st, at selected, st+1 ~ P(s|st,at), rt, at+1 ~ π(a|st+1), st+2 ~ P(s|st+1,at+1), rt+1, …
Q(at|st) = rt + γ.rt+1 + γ2.rt+2 + … = Σk=0∞ γk.rt+k
Same equation as state-value function Q(st) except the action at is not a random
variable but an arbitrary selection.

rt(st,at,st+1) rt+1(st+2,at+1,st+1)

P(st+1|st,at) π(at+1|st+1) P(st+2|st+1,at+1)

st at st+1 at+1 st+2 …

Selected action at, not

drawn from π(a|s).
Xavier Bresson 34
 35

Action-Value Function

Let us compute the value function Q of taking action at in state st considering all possible
episodes drawn from π and P :
et = st, at selected, st+1 ~ P(s|st,at), rt, at+1 ~ π(a|st+1), st+2 ~ P(s|st+1,at+1), rt+1, …
Q(at|st) = !π,P ( rt + γ.rt+1 + γ2.rt+2 + … )

rt(st,at,st+1) rt+1(st+2,at+1,st+1)

P(st+1|st,at) π(at+1|st+1) P(st+2|st+1,at+1)

st at st+1 at+1 st+2 …

Selected action at, not

drawn from π(a|s).

Xavier Bresson 35
 36

Action-Value Function

Let us write the Bellman equation of the action-value function Q :

Q(at|st) = !π,P ( rt + γ.rt+1 + γ2.rt+2 + … )
= !π,P ( rt + !π,P ( γ.rt+1 + γ2.rt+2 + … ) )
= !π,P ( rt + γ. !π,P ( rt+1 + γ.rt+2 + … ) )
= !π,P ( rt + γ. Q(at+1|st+1) )
The sum of total reward Q for taking action at in state st can be decomposed into 2 parts :
The immediate reward rt we get from action at in state st.
The discounted expected sum of future rewards γ.Q(at+1|st+1) we will receive.
It is the expectation over all possible actions at+1 in all possible states st+1.

Xavier Bresson 36
 37

Action-Value Function

The Bellman equation of the action-value function is stationary, i.e. independent of the time step t :
Let st, at, st+1, rt , at+1 be re-written as
s, a, s’, r, a’
Let Q(at|st) = !π,P ( rt + γ. Q(at+1|st+1) ) be re-written as
Q(a|s) = !π,P ( r(s,a,s’)+ γ. Q(a’|s’) )
= Σa’,s’ P(s’|a,s) ( r(s,a,s’) + γ. Q(a’|s’) )
= Σs’ P(s’|a,s) ( r(s,a,s’) + γ. Σa’ Q(a’|s’) )
= !P ( r(s,a,s’) + γ. Σa’ Q(a’|s’) ) r(s,a,s’) a’

There is no more π(a|s) as action a is selected. π(a’|s’)

P(s’|s,a) s’
a’

…
π(a’|s’)

s a r(s,a,s’) a’

π(a’|s’)
s’
Action a is
a’
selected π(a’|s’)

Xavier Bresson 37
 38

Outline

RL and Deep Learning

Agent, Environment and MDP
Policy, Value Function and Model
Optimal Value Function and Policy
Deep Q-Learning (DQN)
Policy Networks
Actor-Critic Algorithm
RL and Supervised Learning
Learning and Planning

Xavier Bresson 38
 39

Optimal Value Function and Policy

Goals of RL :
Find the optimal value function Q*(a|s) :
We search for the maximum reward given by action a in state s.
Find the optimal policy π*(a|s) :
We search for the policy that provides the best action a in state s.
The optimal policy is deterministic, π*(a|s) := π*(s).
Both objectives are equivalent :
Q*(a|s) = maxπ Q(a|s) = maxa Q(a|s) = Qπ*(a|s)
Once the optimal state-value function Q* is found, then the problem is solved
as it is possible to act optimally :
π*(s) = argmaxa Q*(a|s)

Xavier Bresson 39
 40

Optimal Value Function and Policy

Optimal value function Q*(a|s) and optimal policy π*(s) :

Q*(a|s) = maxπ Q(a|s) = maxa Q(a|s) = Qπ*(a|s)
π*(s) = π*(a|s) = argmaxa Q*(a|s)

r(s,a,s’) π*(s’)
a’

π(a’|s’) Q*(a’|s’)
P(s’|s,a) s’
a’
π(a’|s’)
Q(a|s)
a r(s,a,s’)
π*(s) a’
Q*(a|s)
π(a’|s’) Q(a|s)
s’

s
π*(s’)
a’
…
Q*(a’|s’)

Q(a|s)

Xavier Bresson 40
 41

Optimal Value Function and Policy

The Bellman equation of the optimal action-value function (the maximum Q value for taking
action a in state s) :
Q*(a|s) = maxπ Q(a|s)
= maxπ !π,P ( r(s,a,s’)+ γ. Q(a’|s’) )
= maxπ !P ( r(s,a,s’)+ γ. Q(a’|s’) )
= maxπ Σs’ P(s’|a,s) ( r(s,a,s’) + γ. Q(a’|s’) )
r(s,a,s’) π*(s’)
= Σs’ P(s’|a,s) ( r(s,a,s’) + γ. maxπ Q(a’|s’) ) a’

Q*(a’|s’)
= Σs’ P(s’|a,s) ( r(s,a,s’) + γ. maxa’ Q(a’|s’) ) P(s’|s,a) s’
π(a’|s’)

a’
= !s’ ( r + γ. Q*(a’|s’) ) π(a’|s’)
Q(a|s)
a r(s,a,s’)
π*(s) a’
Q*(a|s)
π(a’|s’) Q(a|s)
s’

s
π*(s’)
a’
…
Q*(a’|s’)

Q(a|s)
Xavier Bresson 41
 42

Outline

RL and Deep Learning

Agent, Environment and MDP
Policy, Value Function and Model
Optimal Value Function and Policy
Deep Q-Learning (DQN)
Policy Networks
Actor-Critic Algorithm
RL and Supervised Learning
Learning and Planning

Xavier Bresson 42
 43

Deep Q-Learning

The goal is to estimate the optimal action value function Q* defined as :

Q*(a|s) = !s’ ( r + γ. maxa’ Q(a’|s’) )
It is intractable to compute Q* for all actions and all states.
Q* will be estimated with a neural network (DQN) :
Q*w(a|s) ≈ Q*(a|s) = !s’ ( r + γ. maxa’ Q(a’|s’) )

Target to regress
(supervised learning)

Qw*(a1|s) ≈ Q*(a1|s)
DQN
s

…
Qw*(a|s)
Qw*(aK|s) ≈ Q*(aK|s)

Xavier Bresson 43
 44

Deep Q-Learning

We need to estimate :
Q*(a|s) = !s’ ( r + γ. maxa’ Q(a’|s’) )
= Σs’ P(s’|a,s) ( r(s,a,s’) + γ. maxa’ Q(a’|s’) )
We approximate the analytical optimal value Q* :
Q*(a|s) = !s’ ( r + γ. maxa’ Q(a’|s’) )
≈ 1/|s’| Σs’ ( r(s,a,s’) + γ. maxa’ Q(a’|s’) )
≈ r(s,a,s’) + γ. maxa’ Q(a’|s’) (if only one state s’ is collected)
≈ r(s,a,s’) + γ. maxa’ Q*(a’|s’) (we have maxa’ Q = Q* = maxa’ Q*)

Xavier Bresson 44
 45

Deep Q-Learning

We approximate with DQN the expression :

Q*(a|s) ≈ r(s,a,s’) + γ. maxa’ Q*(a’|s’) ⇒ Q*w(a|s) ≈ r(s,a,s’) + γ. maxa’ Q*w(a’|s’)

Training data (s,a,s’,r) will be collected by rolling out episodes.

DQN MSE loss obtained after rolling out a few of episodes:

L(w) = 1/N Σ(s,a,s’,r) || r(s,a,s’) + γ. maxa’ Q*w(a’|s’) - Q*w(a|s) ||22

where N=|(s,a,s’,r)| is the total number of collected tuples (s,a,s’,r).

Xavier Bresson 45
 46

Vanilla DQN Algorithm

Vanilla DQN algorithm :

Initialize w randomly
Repeat :
Roll-out an episode : st, at ~ πw, rt, st+1
Update state-value function Q*w by gradient descent :
wk+1 = wk – lr . ∇w ( 1/N Σt || r(st,at,st+1) + γ. maxa_t+1 Q*w(at+1|st+1) - Q*w(at|st) ||22 )

DQN does not converge !

Strong correlations between samples.
Predictive function over-fits (memorization) rather than generalizes.
Target values are non-stationary.
The target values change too quickly.

Xavier Bresson 46
 47

Vanilla DQN Algorithm

To remove correlations between training data :

s 1 , a 1 , s 2 , r1
Memory/Experience Replay : s 2 , a 2 , s 3 , r2
Data are randomly sampled from a lookup table. s 3 , a 3 , s 4 , r3
…

st, at, st+1, rt

s, a, s’, r

To stabilize the learning process w.r.t. the signal non-stationarity :

A baseline network wBL is introduced.
This network keeps a copy of the network w for a fixed number of iterations.

Xavier Bresson 47
 48

DQN Algorithm

DQN algorithm :
Initialize w randomly and wBL = w
Repeat :
Roll-out an episode : st, at ~ πw, rt, st+1 and store data in replay memory.
Update state-value function Q*w by mini-batch gradient descent
(batch of randomly sampled data from the lookup table) :
wk+1 = wk – lr . ∇w ( 1/N Σ(s,a,s’,r) || r(s,a,s’) + γ. maxa’ Q*w_BL(a’|s’) - Q*w(a|s) ||22 )
Update baseline network wBL = w if greedy evaluation of network w is better than wBL.

Xavier Bresson 48
 49

DQN Applications

DQN application to Atari games :

Observation ot Action at

Observation is a stack of 4 Agent

consecutive image frames
of the game.
Action is the joystick
position and a button.
Observation ot+1
Reward rt
Environment

Reward is a score
for this action
and state.

Xavier Bresson 49
 50

DQN Applications

DQN application to Atari games :

Learning of values Qw(a|s) from input s.
Input state s is a stack of raw pixels from last 4 frames.
Output is Qw(a|s) for the 18 joystick/button positions.

Xavier Bresson 50
 51

DQN Applications

Application to Atari games :

Xavier Bresson 51
 52

DQN Applications

Learning optimal strategy :

Xavier Bresson 52
 53

DQN Applications

Application to Mario Bros :

https://github.com/aleju/mario-ai

Xavier Bresson 53
 54

Demo 01
Finding optimal strategy to stabilize a pole attached by an un-actuated joint to a cart,
which moves along a frictionless track.
The pendulum starts upright, and the goal is to prevent it from falling over by
increasing and reducing the cart's velocity.
The cart-pole environment is the “MNIST” experiment for RL.
Use it to develop new RL architectures !

Optimal strategy
Xavier Bresson 54
 55

Demo 01

Replay Memory

Xavier Bresson 55
 56

Q network Demo 01

Compute Q
scores given a
state

Select action
given a state

Select random action

to explore the space
of (state,action)
Loss

Read (s,a,s’,r) in
memory

Standard
MSE loss

Xavier Bresson 56
 57

Demo 01

Rollout class

Rollout one
episode

Write (s,a,s’,r)
in memory

Xavier Bresson 57
 58

Demo 01

Rollout one
episode.
Compute loss,
gradient, update.

Xavier Bresson 58
 59
Train multiple epochs Demo 01 Train one epoch
by rolling out a
few episodes

Smoothly decreasing the

random action prob
(space exploration
parameter)

Updating the
baseline network if
greedy evaluation
of current Q
network is better.

Stopping condition
(high reward)

Xavier Bresson 59
 60

DQN Algorithm

Advantages :
Memory Replay module memorizes the history of (state,action,reward).
Examples : All professional games of Go, Chess, etc can be memorized.
Memory Replay is a nice attractive feature toward AI.
Challenges :
Hard to compute a good baseline.
Hard to control the exploration of the space (state,action).

Xavier Bresson 60
 61

Outline

RL and Deep Learning

Agent, Environment and MDP
Policy, Value Function and Model
Optimal Value Function and Policy
Deep Q-Learning (DQN)
Policy Networks
Actor-Critic Algorithm
RL and Supervised Learning
Learning and Planning

Xavier Bresson 61
 62

Policy Networks
Find optimal policy that maximizes an expected reward :
Local discarded rewards :
Optimize the expected reward of all actions and states :
maxπ !a,s ( Q(a|s) )
Q(a|s) : Local discarded rewards for taking action a in state s in e.g. cartpole
simulation.
Global/sparse rewards :
Optimize the expected reward of all episodes/trajectories :
maxπ !e ~ π,P ( Q(e) )
Q(e) : Global rewards s.a. win/loss in Go or length of cartpole episode.

Xavier Bresson 62
 63

Policy Network for Local Discarded Rewards

Optimize the expected reward of all actions and states : π(a|s) a

maxπ !a,s ( Q(a|s) ) s

Σa,s π(a|s) . Q(a|s)

π(a|s)
a

It is intractable to compute the exact policy function π in high-dimensional spaces.

We will approximate the policy function π with a neural network :

π(a|s) ⇒ πw(a|s)

Xavier Bresson 63
 64

Policy Gradient
Policy gradient :
∇w ( "a,s ( Q(a|s) ) = ∇w ( Σa,s πw(a|s) . Q(a|s) )
= Σa,s ( ∇w πw(a|s) ) . Q(a|s) )
= Σa,s πw(a|s)/πw(a|s) ∇w πw(a|s) . Q(a|s)
= Σa,s πw(a|s) ( ∇w πw(a|s)/πw(a|s) ) . Q(a|s)
= Σa,s πw(a|s) ∇w log πw(a|s) . Q(a|s)
= "a,s ( ∇w log πw(a|s) . Q(a|s) )
= ∇w ( "a,s ( log πw(a|s) . Q(a|s) ) )

+
Policy loss for local discarded rewards :

LL(w) = "a,s ( - log πw(a|s) . Q(a|s) )

minw LL(w)

Xavier Bresson 64
 65

Interpretation of Policy Gradient

Policy gradient :
!a,s ( - ∇w log πw(a|s) . Q(a|s) )

Score function Reward for action a

(log likelihood) in state s

Gradient of score function favors actions

of high probability, and diminishes
actions with low probability.

This term increases probability of actions

with positive reward Q, and decreases
probability of actions with negative reward Q.

Xavier Bresson 65
 66

Policy Gradient Approximation

It is intractable to compute the exact gradient in high-dimensional spaces.

We will approximate it by running multiple episodes.
One episode roll-out :
e = s1, a1 ~ π(a|s1), s2 ~ P(s|s1,a1), r1, a2 ~ π(a|s2), s3 ~ P(s|s2,a2), r2,
… sT ~ P(s|sT-1,aT-1), aT ~ π(a|sT), sT+1 ~ P(s|sT,aT), rT

r1(s1,a1,s2) r2(s2,a2,s3) rT(sT,aT,sT+1)

π(a1|s1) P(s2|s1,a1) π(a2|s2) P(s3|s2,a2) π(aT|sT) P(sT+1|sT,aT)

s1 a1 s2 a2 s3 … aT sT+1

Xavier Bresson 66
 67

Loss Approximation
Loss for one episode :
LL(w) = !a,s ( - log πw(a|s) . Q(a|s) )
≈ 1/T Σt - log πw(at|st) . Q(at|st)

Approximation of Number of time

expected value steps in one episode
for one episode

with Q(at|st) = rt + γ.rt+1 + γ2.rt+2 + … + γT.rT Discounted reward

Loss for a mini-batch of M episodes :

LL(w) ≈ 1/M Σm 1/Tm Σt - log πw(atm|stm) . Q(atm|stm)

Xavier Bresson 67
 68

REINFORCE Algorithm

Deep policy network algorithm :

Initialize the policy network at random wk=0 for πw(a|s).

Repeat :

Collect a mini-batch of M episodes em by rolling out the current policy network πw.

Update the policy network by stochastic gradient descent :

wk+1 = wk – lr . ∇w ( 1/M Σm 1/Tm Σt - log πw(atm|stm) . Q(atm|stm) )

≈ ∇w LL(w)

This algorithm is known as REINFORCE or Monte-Carlo policy gradient.

Xavier Bresson 68
 69

Demo 02

Finding the optimal policy network with local discarded rewards.

Optimal policy network

Xavier Bresson 69
 70

Policy network Demo 02

Compute action
probabilities
given a state s

Selection an
action a given a
state s
Select action by
Bernouilli random
sampling to explore Compute
the space of discarded
(state,action) rewards
Loss
Z-score
discarded
rewards

Policy loss for

one episode
Xavier Bresson 70
 71

Demo 02

Rollout class

Rollout one
episode

Collect all local

rewards and log
probability of
actions

Xavier Bresson 71
 72

Demo 02

Rollout one
episode.
Compute loss,
gradient, update.

Xavier Bresson 72
 73

Demo 02
Train one epoch
Train multiple epochs
by rolling out a
few episodes

Stopping condition
(high reward)

Xavier Bresson 73
 74

Policy Network for Global Rewards

Optimize the expected reward of all episodes/trajectories :

maxπ !e ~ π,P ( Q(e) ) st+1

Σe Pe(e|π,P) . Q(e) s3
at
st+1
at
a2 st+1
s3
s2 at
a2 st+1
s3
a1 at
s2 a2 st+1
s3
Probability of s1 a1 s2
…
at
st+1
a2 s3
episode e given π,P at
st+1
s2 a2 s3
a1 at
π(a1|s1) st+1

a2 s3
s2 at
st+1
P(s2|s1,a1) s3
a2 at
π(a2|s2) st+1
s3 at
P(s3|s2,a2) st+1
P(st+1|st,at)

Let us represent the policy function π with a neural network :

π(a|s) ⇒ πw(a|s)

Xavier Bresson 74
 75

Policy Gradient
Policy gradient :
∇w ( "e ~ π,P ( Q(e) ) ) = ∇w ( Σe Pe(e|πw,P) . Q(e) )
= Σe ∇w Pe(e|πw,P) . Q(e)
= Σe Pe(e|πw,P) ( ∇w Pe(e|πw,P)/Pe(e|πw,P) ) . Q(e)
= Σe Pe(e|πw,P) ∇w log Pe(e|πw,P) . Q(e)
= "e ( ∇w log Pe(e|πw,P) . Q(e) )
≈ 1/M Σm ∇w log Pe(em|πw,P) . Q(em)
Approximation of
expected value
with a mini-batch Probability of having the mth episode
of M episodes
Pe(em|πw,P) = ∏t πw(atm |stm ) . P(st+1m |stm,atm)

Xavier Bresson 75
 76

Policy Gradient
Policy gradient :
∇w ( "e ~ π,P ( Q(e) ) ) ≈ 1/M Σm ∇w log Pe(em|πw,P) . Q(em)
≈ 1/M Σm ∇w log ( ∏t πw(atm |stm ) . P(st+1m |stm,atm) ) . Q(em)
≈ 1/M Σm ∇w ( Σt log πw(atm |stm ) + Σt log P(st+1m |stm,atm) ) . Q(em)

Gradient is zero as term is independent of w.

The policy gradient is independent of the
dynamics of the environment.
No need to know the dynamics of the world
to act optimally !

∇w ( "e ~ π,P ( Q(e) ) ) ≈ ∇w ( 1/M Σm Σt log πw(atm |stm ) . Q(em) )

+
Policy loss for global reward :
LG(w) = 1/M Σm Σt - log πw(atm |stm ) . Q(em)
minw LG(w)
Xavier Bresson 76
 77

Demo 03

Finding the optimal policy network with global rewards.

Optimal policy network

Xavier Bresson 77
 78

Policy network
Demo 03

Compute action
probabilities
given a state s

Selection an
action a given a
state s

Loss

Compare
current policy
with baseline

Policy loss for

one episode
Xavier Bresson 78
 79

Demo 03
Rollout class

Rollout one
episode

Collect all global

rewards, i.e.
length of each
episode

Xavier Bresson 79
 80

Demo 03

Rollout one
episode for current
and baseline
policies.
Compute loss,
gradient, update.

Xavier Bresson 80
 81

Train multiple epochs

Demo 03 Train one epoch
by rolling out a
few episodes

Updating the baseline

policy if greedy
evaluation of current
policy is better.

Stopping condition
(high reward)

Xavier Bresson 81
 82

Policy Algorithms

Policy network with local rewards :

Good to receive lots of local rewards to learn faster and better.
Hard to define a good baseline (use z-score of local rewards of mini-batch of episodes).
Slow to learn long-term strategy

Policy network with global rewards :

We receive only a global/sparse reward at the end of the episode.
Easy to define a good baseline (by greedy roll-out)
Faster to learn long-term strategy.

Xavier Bresson 82
 83

Outline

RL and Deep Learning

Agent, Environment and MDP
Policy, Value Function and Model
Optimal Value Function and Policy
Deep Q-Learning (DQN)
Policy Networks
Actor-Critic Algorithm
RL and Supervised Learning
Learning and Planning

Xavier Bresson 83
 84

Actor-Critic Algorithm

Monte-Carlo policy gradient (i.e. policy w/ local discarded rewards) a.k.a. REINFORCE has
a high variance (because of Q(at|st)) :
∇w LL(w) = Σt - ∇w log πw(at|st) . Q(at|st)
where Q(at|st) = rt + γ.rt+1 + γ2.rt+2 + … + γT.rT = Σk=tT γk-t.rk
The variance of the policy (actor) can be reduced by constraining Q(at|st) (critic) to satisfy
the Bellman equation :
Q(at|st) = "π,P ( rt + γ. Q(at+1|st+1) )
≈ rt + γ. Q(at+1|st+1) (during one episode)
Q(a|s) will be estimated with a neural network.
One-Step Actor-Critic (QAC) loss :
LQAC(w,v) = Σt - log πw(at|st) . Qv(at|st)
+ Σt || Qv(at|st) – ( rt + γ. Qv(at+1|st+1) ) ||22

Xavier Bresson 84
 85

QAC Algorithm

One-Step Actor-Critic (QAC) algorithm :

Initialize w and v randomly.
Repeat :
Roll-out : st, at ~ πw, rt, st+1, at+1 ~ πw
Error : et = Qv(at|st) – ( rt + γ. Qv(at+1|st+1) )
Actor update : wk+1 = wk - lr . Σt ∇w - log πw(at|st) . Qv(at|st)
Critic update : vk+1 = vk - lr . et . ∇v Qv(at|st)
We simultaneously estimate the policy πw (actor) and the action-value function Qv (critic)
that evaluates the policy :
Qv(at|st) is the evaluation of long-term returned by the critic at time t.
et is the estimated error of long-term returned at time t.

Xavier Bresson 85
 86

Demo 04

Xavier Bresson 86
 87

AAC Algorithm

One-Step Actor-Critic (QAC) loss :

LQAC(w,v) = Σt - log πw(at|st) . Qv(at|st)
+ Σt || Qv(at|st) – ( rt + γ. Qv(at+1|st+1) ) ||22
and its policy gradient :
∇w LP(w) = Σt - ∇w log πw(at|st) . Qv(at|st)
is biased and may not find the optimal policy.
A baseline function b(s) is introduced to remove the bias :
∇w LP(w) = Σt - ∇w log πw(at|st) . ( Qv(at|st) – b(st) )
A good baseline is
b(st) = Q(st), the state-value function in state st, i.e. the future reward from state st.
Q(s) will be estimated with a neural network.

Xavier Bresson 87
 88

AAC Algorithm

Advantage Actor-Critic (AAC or A2C) loss :

LAAC(w,v,u) = Σt - log πw(at|st) . ( Qv(at|st) – Qu(st) )
+ Σt || Qv(at|st) – ( rt + γ. Qv(at+1|st+1) ) ||22
+ Σt || Qu(st) - Q(at|st) ||22 with Q(at|st) = Σk=tT γk-t.rk
and its policy gradient :
∇w LAAC(w) = Σt - ∇w log πw(at|st) . A(at|st)
where A(at|st) = Qv(at|st) – Qu(st) is the Advantage function.
This gradient policy is unbiased and has reduced variance.

Xavier Bresson 88
 89

AAC Algorithm

Advantage Actor-Critic (AAC or A2C) algorithm :

Initialize w, v, u randomly.
Repeat :
Roll-out an episode and write in memory : st, at ~ πw, rt, st+1, at+1 ~ πw
Discounted reward : Q(at|st) = Σk=tT γk-t.rk
Errors : et1 = Qv(at|st) – ( rt + γ. Qv(at+1|st+1) ) and et2 = Qu(st) - Q(at|st)
Advantage function : A(at|st) = Qv(at|st) – Qu(st)
Update : wk+1 = wk - lr . Σt ∇w - log πw(at|st) . A(at|st)
Update : vk+1 = vk - lr . et1 . ∇v Qv(at|st)
Update : uk+1 = uk - lr . et2 . ∇v Qu(st)

Xavier Bresson 89
 90

Actor-critic network Demo 05

Policy (actor)
Evaluate action-
value Q function
(critic)

Evaluate state-
value Q function
(critic)

Action selected
by policy

Xavier Bresson 90
 91

Demo 05
Loss
Reward must be
negative

Policy loss for

one episode

State-value Q
function loss for
one episode

State Q function
loss for one
episode

Xavier Bresson 91
 92

Demo 05
Rollout class

Rollout one
episode

Write in memory
(s,a,s’,r)

Xavier Bresson 92
 93

Demo 05

Rollout one
episode for current
and baseline
policy.
Compute loss,
gradient, update.

Xavier Bresson 93
 94

Demo 05
Train multiple epochs Train one epoch
by rolling out a
few episodes

Stopping condition
(high reward)

Xavier Bresson 94
 95

Continuous Actor-Critic Algorithm

Continuous control :
Deterministic policy and continuous action space :
a = π(s)
a in R (continuous variable)
Find the deterministic policy π that maximizes the expected total reward over all possible
actions and states:
maxπ !a=π(s),s ( Q(a|s) )
Σa=π(s),s Q(a|s)
Σs Q(π(s)|s)

Xavier Bresson 95
 96

Continuous Actor-Critic Algorithm

Continuous control :
Policy gradient :
∇w ( "a=π(s),s ( Q(a|s) ) = ∇w ( "s ( Q(πw(s)|s) )
= "s ( ∇w ( Q(πw(s)|s) )
= "s ( dQ(πw)/dπw . ∇w πw ) // chain rule on differentiable Q
(as πw=a in R)
= "s ( dQ(a|s)/da . ∇w a ) // as πw=a

Gradient of the Q-value function. How to adjust the policy so

How to adjust the Q-value to take action that we can take these actions
a in state s more often (or less often). more often (or less often).

The product is the gradient to follow to make

our policy better to get higher Q value.

Xavier Bresson 96
 97

Further Improvements

Observations ot are raw pixels from current frame.

Observations ot of the maze are partial.
State st =f(o1,…,ot) is modeled by a recurrent neural network (LSTM).
One network to output action/state functions Q(a|s), Q(s) and policy function π(a|s).
Task is to collect apples (+1 reward) and escape (+10 reward).

Xavier Bresson 97
 98

Applications

Learning locomotion Learning human-like robot hand gesture

Humanoid and spider Application to Rubik’s Cube

Xavier Bresson 98
 99

Outline

RL and Deep Learning

Agent, Environment and MDP
Policy, Value Function and Model
Optimal Value Function and Policy
Deep Q-Learning (DQN)
Policy Networks
Actor-Critic Algorithm
RL and Supervised Learning
Learning and Planning

Xavier Bresson 99
 100

Supervised Learning

SL is learning with a teacher who tells the best action to take given any state of the environment.
Training set for SL :
(st,at)t=1:T where at is the label/target action that optimizes the return.
By analogy to computer vision :
s = image
a = class
Likelihood probability of class a given image s :
π(a|s)
It corresponds to the policy function π in RL.

Xavier Bresson 100

 101

Supervised Learning

Loss function :
minw Lπ(W) = Σt cross_entropy ( πw(at|st) , π*(at|st) )
= - Σt log πw(at|st) . π*(at|st)
maxw Lπ(W) = - minw Lπ(W) πw(a|s)

= Σt log πw(at|st) . π*(at|st)

Gradient of loss function : a1 a2 a* aK

∇w Lπ(W) = ∇w ( Σt log πw(at|st) . π*(at|st) ) Predicted probability

= Σt ∇w log πw(at|st) . π*(at|st) π*(a|s)

1
In SL, we know the best
action to take to optimize
the reward function (i.e. the
a1 a2 a* aK
cross entropy loss).
Target probability

Xavier Bresson 101

 102

Reinforcement Learning

RL has no teacher (no label) to tell the best action to take.

RL is a trial-and-error learning paradigm.
RL draws sequence of actions that provide reward.
Depending on the reward, the actions are encouraged and discouraged.
Training set :
A batch of episodes drawn from probabilities π and P :
eT = s1, a1 ~ π(a|s1), s2 ~ P(s|s1,a1), r1, a2 ~ π(a|s2), s3 ~ P(s|s2,a2), r2,
… sT ~ P(s|sT-1,aT-1), aT ~ π(a|sT), sT+1 ~ P(s|sT,aT), rT
For each state and action, a reward rt is arbitrary defined :
Game Go : rt=0,1,-1 (1 for won game, -1 for lost game, 0 otherwise).
Robot : 1 if it does not fall, -1 if it falls.
Reward can be highly sparse, unlike SL.

Xavier Bresson 102

 103

RL and SL Similarities

Gradient of loss function :

Reinforcement learning :
∇w R(πw) = "a,s ( ∇w log πw(a|s) . Q(a|s) )
= Σt ∇w log πw(at|st) . Q(at|st)

In RL, we estimate the future

reward of taking action at in
state st.

Supervised learning :
∇w Lπ(w) = ∇w ( Σt log πw(at|st) . π*(at|st) )
= Σt ∇w log πw(at|st) . π*(at|st)

In SL, we know the best

action at to take in state st.

Xavier Bresson 103

 104

Outline

RL and Deep Learning

Agent, Environment and MDP
Policy, Value Function and Model
Optimal Value Function and Policy
Deep Q-Learning (DQN)
Policy Networks
Actor-Critic Algorithm
RL and Supervised Learning
Learning and Planning

Xavier Bresson 104

 105

Learning and Planning

RL :
Agent interacts with the environment to understand it and improve its policy to get the
best possible reward from this environment.
Agent learns implicitly the environment.

Planning :
We assume the environment is known (or it was learned).
The agent can inquiry multiple times the simulated environment (no real action taken in
the real environment), and decides what to do from these inquiries.
The agent is “thinking” what to do before acting.

RL and planning are usually done together.

Xavier Bresson 105

 106

Learning and Planning

If the environment is high-dimensional :

Curse of dimensionality :
The space is too large to simultaneously explore and exploit.
How to balance the trade-off between exploration and exploitation ?
Exploration :
Give up some rewards to get more information about the environment, which can
eventually lead to higher rewards.
Exploitation :
Give up potential more rewards by only using the known environment to get good rewards.

Xavier Bresson 106

 107

Learning and Planning

Examples :
Game of Go :
Play the move that is known to be the best.
Try a new strategy.
Online advertisement :
Show the consumer the most successful ads.
Try a new ad.

Xavier Bresson 107

 108

Learning and Planning

Model-based RL :
Learn a model of a proxy of the environment.
This model can be used to interact virtually with the environment, just as we would
interact with the real environment.
A model is essential to do a look-ahead search by querying the environment with some
actions and observe the consequences/rewards of this action.
This allows to think ahead and make better actions in high-dimensional spaces.

Xavier Bresson 108

 109

Questions?

Xavier Bresson 109