Machine Learning

Machine Learning

Lecture 11: Reinforcement Learning
(thanks in part to Bill Smart at Washington University in St. Louis)

From Bryan Pardo, Northwestern University EECS 349

Learning Types

• Supervised learning:
– (Input, output) pairs of the function to be learned can
be perceived or are given.
Back-propagation in Neural Nets

• Unsupervised Learning:
– No information about desired outcomes given

K-means clustering

• Reinforcement learning:
– Reward or punishment for actions

Q-Learning


Reinforcement Learning

• Task
– Learn how to behave to achieve a goal
– Learn through experience from trial and error

• Examples
– Game playing: The agent knows when it wins, but
doesn’t know the appropriate action in each state
along the way

– Control: a traffic system can measure the delay of
cars, but not know how to decrease it.


Basic RL Model

1. Observe state, st
2. Decide on an action, at World

3. Perform action
4. Observe new state, st+1 S R A
5. Observe reward, rt+1
6. Learn from experience
7. Repeat

•Goal: Find a control policy that will maximize the
observed rewards over the lifetime of the agent

An Example: Gridworld

• Canonical RL domain
States are grid cells +1

4 actions: N, S, E, W
Reward for entering top right cell
-0.01 for every other move


Mathematics of RL

• Before we talk about RL, we need to cover
some background material
– Simple decision theory
– Markov Decision Processes
– Value functions
– Dynamic programming


Making Single Decisions
• Single decision to be made
– Multiple discrete actions
– Each action has a reward associated with it
• Goal is to maximize reward
– Not hard: just pick the action with the largest reward
• State 0 has a value of 2
– Sum of rewards from taking the best action from the
state
1
1
0
2
2

Markov Decision Processes

• We can generalize the previous example
to multiple sequential decisions
– Each decision affects subsequent decisions

• This is formally modeled by a Markov
Decision Process (MDP)
A 3
1 A
1 1 1
A B
0 1 5
B
2 -1000 10
2 A 4 A


Markov Decision Processes

• Formally, a MDP is
– A set of states, S = {s1, s2, ... , sn}
– A set of actions, A = {a1, a2, ... , am}
– A reward function, R: S A S→
– A transition function, Pija P st 1 j | st i, at a
• Sometimes T: S A→S (and thus R: S A→ )

• We want to learn a policy, : S →A
– Maximize sum of rewards we see over our
lifetime


Policies
• A policy (s) returns what action to take
in state s.
• There are 3 policies for this MDP
Policy 1: 0 →1 →3 →5
Policy 2: 0 →1 →4 →5
Policy 3: 0 →2 →4 →5
A 3
1 A
1 1 1
A B
0 1 5
B
2 -1000 10
2 A 4 A


Comparing Policies

• Which policy is best?
• Order them by how much reward they see

Policy 1: 0 →1 →3 →5 = 1 + 1 + 1 = 3
Policy 2: 0 →1 →4 →5 = 1 + 1 + 10 = 12
Policy 3: 0 →2 →4 →5 = 2 – 1000 + 10 = -988
A 3
1 A
1 1 1
A B
0 1 5
B
2 -1000 10
2 A 4 A


Value Functions
• We can associate a value with each state
– For a fixed policy
– How good is it to run policy from that state s
– This is the state value function, V

V1(s1) = 2
V1(s0) = 3 V2(s1) = 11 V1(s0) = 1
V2(s0) = 12 A
V3(s0) = -988 1 3 A
1 1 1
A B
0 1 5
B
2 -1000 10
2 4 A
A V2(s0) = 10
V3(s0) = -990 V3(s0) = 10

Q Functions

• Define value without specifying the policy
– Specify the value of taking action A from state S and
then performing optimally, thereafter

Q(1, A) = 2 How do you tell which
Q(1, B) = 11 Q(3, A) = 1 action to take from
Q(0, A) = 12 A 3 each state?
Q(0, B) = -988 1 A
1 1 1
A B
0 1 5
B
2 -1000 10
2 A 4 A
Q(2, A) = -990 Q(4, A) = 10


Value Functions
• So, we have two value functions
s’ is the
next state

V (s) = R(s, (s), s’) + V (s’) a’ is the
next action

Q(s, a) = R(s, a, s’) + maxa’ Q(s’, a’)

• Both have the same form
– Next reward plus the best I can do from the next state


Value Functions
• These can be extend to probabilistic actions
(for when the results of an action are not certain, or
when a policy is probabilistic)

V s P s'| s, (s) R s, s , s' V s'
s'

Q s,a P(s'| s,a) R s, a, s' max a' Q s', a'
s'


Getting the Policy

• If we have the value function, then finding
the optimal policy, (s), is easy…just find
the policy that maximizes value

(s) = arg maxa (R(s, a, s’) + V (s’))
(s) = arg maxa Q(s, a)


Problems with Our Functions
• Consider this MDP
– Number of steps is now unlimited because of loops
– Value of states 1 and 2 is infinite for some policies
Q(1, A) = 1 + Q(1, A)
Q(1, A) = 1 + 1 + Q(1, A)
Q(1, A) = 1 + 1 + 1 + Q(1, A) 1
A
Q(1, A) = ...
1
• This is bad -1000
B 0
A
– All policies with a non- 0 3
zero reward cycle have B B 0
infinite value 1000 2
A
1


Better Value Functions
• Introduce the discount factor to get around
the problem of infinite value
– Three interpretations
• Probability of living to see the next time step
• Measure of the uncertainty inherent in the world
• Makes the mathematics work out nicely

Assume 0 ≤ ≤ 1

V (s) = R(s, (s), s’) + V (s’)

Q(s, a) = R(s, a, s’) + maxa’ Q(s’, a’)


Better Value Functions

1
Q(1, A)
1
1 Q(1, B) 0
Q(0, A) 1000 A
1

Q(1, B) 1000 -1000 1 0
1
A B
0 3
• Optimal Policy: B B 0
(0) = B 1000 2
1
(1) = A A Q(2, A)
1 1
(2) = A Q(2, B) 0


Dynamic Programming

• Given the complete MDP model, we can
compute the optimal value function directly

V(1) = 1 + 10 2 V(3) = 1 + 0
A
1 3 A
1 1 1 V(5) = 0
A B
0 1 5 0
B A
V(0) = 1 + + 10 2 +0 3
2 -1000 10
2 4 A
A
V(2) = - 1000 +10 V(4) = 10 + 0

[Bertsekas, 87, 95a, 95b]


Reinforcement Learning
• What happens if we don’t have the whole MDP?
– We know the states and actions
– We don’t have the system model (transition function)
or reward function
• We’re only allowed to sample from the MDP
– Can observe experiences (s, a, r, s’)
– Need to perform actions to generate new experiences
• This is Reinforcement Learning (RL)
– Sometimes called Approximate Dynamic
Programming (ADP)


Learning Value Functions

• We still want to learn a value function
– We’re forced to approximate it iteratively
– Based on direct experience of the world

• Four main algorithms
– Certainty equivalence
– TD learning
– Q-learning
– SARSA


Certainty Equivalence
• Collect experience by moving through the world
– s0, a0, r1, s1, a1, r2, s2, a2, r3, s3, a3, r4, s4, a4, r5, s5, ...

• Use these to estimate the underlying MDP
– Transition function, T: S A → S
– Reward function, R: S A S →

• Compute the optimal value function for this MDP
– And then compute the optimal policy from it


How are we going to do this?
100

G
points • Reward whole
policies?
– That could be a pain

• What about
incremental rewards?
S
– Everything is a 0
except for the goal

• Now what???


Q-Learning [Watkins & Dayan, 92]

• Q-learning iteratively approximates the state-
action value function, Q
– We won’t estimate the MDP directly
– Learns the value function and policy simultaneously

• Keep an estimate of Q(s, a) in a table
– Update these estimates as we gather more
experience
– Estimates do not depend on exploration policy
– Q-learning is an off-policy method


Q-Learning Algorithm

1. Initialize Q(s, a) to small random values, s, a
(often they really use 0 as the starting value)
2. Observe state, s
3. Randomly pick an action, a, and do it
4. Observe next state, s’, and reward, r
5. Q(s, a) ← (1 - )Q(s, a) + (r + maxa’Q(s’, a’))
6. Go to 2

• 0≤ ≤ 1 is the learning rate
– We need to decay this, just like TD

Q-learning
• Q-learning, learns the expected utility of
taking a particular action a in state s

0 100 0 0
90 100 0
0 G G 90 100 G
0 0
72 81 81
0 0 0 0 81 90
100 100
81 90 100 81 90
0 0
72 81

r(state, action) V*(state) values
immediate reward values Q(state, action) values


Exploration vs. Exploitation
• We want to pick good actions most of the
time, but also do some exploration
– Exploring means that we can learn better policies
– But, we want to balance known good actions with
exploratory ones
– This is called the exploration/exploitation problem


Picking Actions
-greedy
– Pick best (greedy) action with probability
– Otherwise, pick a random action
• Boltzmann (Soft-Max)
– Pick an action based on its Q-value
Q(s, a)

e
P(a | s) Q(s, a' )

e
a'
…where is the “temperature”


On-Policy vs. Off Policy
• On-policy algorithms
– Final policy is influenced by the exploration policy
– Generally, the exploration policy needs to be “close”
to the final policy
– Can get stuck in local maxima

• Off-policy algorithms
– Final policy is independent of exploration policy
– Can use arbitrary exploration policies
– Will not get stuck in local maxima


SARSA
• SARSA iteratively approximates the state-action
value function, Q
– Like Q-learning, SARSA learns the policy and the
value function simultaneously

• Keep an estimate of Q(s, a) in a table
– Update these estimates based on experiences
– Estimates depend on the exploration policy
– SARSA is an on-policy method
– Policy is derived from current value estimates


SARSA Algorithm
1. Initialize Q(s, a) to small random values, s, a
2. Observe state, s
3. Pick an action ACCORDING TO A POLICY
4. Observe next state, s’, and reward, r
5. Q(s, a) ← (1- )Q(s, a) + (r + Q(s’, (s’)))
6. Go to 2

• 0≤ ≤ 1 is the learning rate
– We need to decay this, just like TD


TD
• TD-learning estimates the value function directly
[Sutton, 88]
– Don’t try to learn the underlying MDP

• Keep an estimate of V (s) in a table
– Update these estimates as we gather more
experience
– Estimates depend on exploration policy,
– TD is an on-policy method


TD-Learning Algorithm
1. Initialize V (s) to 0, and e(s) = 0 s
2. Observe state, s
3. Perform action according to the policy (s)
4. Observe new state, s’, and reward, r
← r + V (s’) - V (s)
6. e(s) ← e(s)+1
7. For all states j
V (s) ← V (s) + e(j)
e(j) ← e(s) = future returns
discount factor
8. Go to 2 = eligibility discount
= learning rate

TD-Learning
• V (s) is guaranteed to converge to V*(s)
– After an infinite number of experiences
– If we decay the learning rate
2
t t
t 0 t 0

c
t will work
c t

• In practice, we often don’t need value
convergence
– Policy convergence generally happens sooner


Convergence Guarantees
• The convergence guarantees for RL are “in the
limit”
– The word “infinite” crops up several times

• Don’t let this put you off
– Value convergence is different than policy
convergence
– We’re more interested in policy convergence
– If one action is really better than the others, policy
convergence will happen relatively quickly


Rewards
• Rewards measure how well the policy is doing
– Often correspond to events in the world
• Current load on a machine
• Reaching the coffee machine
• Program crashing
– Everything else gets a 0 reward

• Things work better if the rewards are
incremental
– For example, distance to goal at each step
– These reward functions are often hard to design


The Markov Property
• RL needs a set of states that are Markov
– Everything you need to know to make a decision is
included in the state
– Not allowed to consult the past
Not holding key

• Rule-of-thumb Holding key

– If you can calculate the reward K
function from the state without
any additional information,
you’re OK

S G


But, What’s the Catch?

• RL will solve all of your problems, but
– We need lots of experience to train from
– Taking random actions can be dangerous
– It can take a long time to learn
– Not all problems fit into the MDP framework


Learning Policies Directly
• An alternative approach to RL is to reward whole
policies, rather than individual actions
– Run whole policy, then receive a single reward
– Reward measures success of the whole policy

• If there are a small number of policies, we can
exhaustively try them all
– However, this is not possible in most interesting
problems


Policy Gradient Methods
• Assume that our policy, p, has a set of n real-
valued parameters, q = {q1, q2, q3, ... , qn }
– Running the policy with a particular q results in a
reward, rq
R
– Estimate the reward gradient, , for each qi
θi

R
θi θi
θi

This is another
learning rate


Policy Gradient Methods
• This results in hill-climbing in policy space
– So, it’s subject to all the problems of hill-climbing
– But, we can also use tricks from search, like random
restarts and momentum terms

• This is a good approach if you have a
parameterized policy
– Typically faster than value-based methods
– “Safe” exploration, if you have a good policy
– Learns locally-best parameters for that policy


An Example: Learning to Walk
[Kohl & Stone, 04]

• RoboCup legged league
– Walking quickly is a big advantage

• Robots have a parameterized gait controller
– 11 parameters
– Controls step length, height, etc.

• Robots walk across soccer pitch and are timed
– Reward is a function of the time taken



• Basic idea
1. Pick an initial = { 1, 2, ... , 11}
2. Generate N testing parameter settings by perturbing
j ={ +
1 1, 2 + 2, ... , 11 + 11}, i {- , 0, }
3. Test each setting, and observe rewards
j→ rj
4. For each i if θi largest
0
Calculate 1
+,
1
0,
1
- and set θ'i θi 0 if θi largest
5. Set ← ’, and go to 2 if θi largest

Average reward
when qni = qi - di


Initial Final

http://utopia.utexas.edu/media/features/av.qtl

Video: Nate Kohl & Peter Stone, UT Austin


Value Function or Policy Gradient?

• When should I use policy gradient?
– When there’s a parameterized policy
– When there’s a high-dimensional state space
– When we expect the gradient to be smooth

• When should I use a value-based
method?
– When there is no parameterized policy
– When we have no idea how to solve the
problem

Machine Learning

More Related Content

Similar to Machine Learning (20)

More from butest (20)

Machine Learning