Multi-armed Bandits

Multi-armed Bandits
Dongmin Lee
RLI Study

Reference
- Reinforcement Learning : An Introduction,
Richard S. Sutton & Andrew G. Barto
(Link)
- 멀티 암드 밴딧(Multi-Armed Bandits), 송호연
(https://brunch.co.kr/@chris-song/62)

Index
1. Why do you need to know Multi-armed Bandits?
2. A k-armed Bandit Problem
3. Simple-average Action-value Methods
4. A simple Bandit Algorithm
5. Gradient Bandit Algorithm
6. Summary

1. Why do you need to know
Multi-armed Bandits(MAB)?

1. Why do you need to know MAB?
- Reinforcement Learning(RL) uses training information that
‘Evaluate’(‘Instruct’ X) the actions
- Evaluative feedback indicates ‘How good the action taken was’
- Because of simplicity, ‘Nonassociative’ one situation
- Most prior work involving evaluative feedback
- ‘Nonassociative’, ‘Evaluative feedback’ -> MAB
- In order to Introduce basic learning methods in later chapters

1. Why do you need to know MAB?
In my opinion,
- I think we can’t seem to know RL without knowing MAB.
- MAB deal with ‘Exploitation & Exploration’ of the core ideas in RL.
- In the full reinforcement learning problem, MAB is always used.
- In every profession, MAB is very useful.

Do you know what MAB is?

Source : Multi-Armed Bandit
Image source : https://brunch.co.kr/@chris-song/62
- Slot Machine -> Bandit
- Slot Machine’s lever -> Armed
- N slot Machine
 Multi-armed Bandits

Among the various slot machines,
which slot machine
should I put my money on
and lower the lever?

How can you make the best return
on your investment?

MAB is a algorithm created to
optimize investment in slot machines

A K-armed Bandit Problem

Source : Reinforcement Learning : An Introduction, Richard S. Sutton & Andrew G. Barto
Image source : https://drive.google.com/file/d/1xeUDVGWGUUv1-ccUMAZHJLej2C7aAFWY/view
t – Discrete time step or play number
𝐴 𝑡 - Action at time t
𝑅𝑡 - Reward at time t
𝑞∗ 𝑎 – True value (expected reward) of action a

In our k-armed bandit problem, each of the k actions has
an expected or mean reward given that that action is selected;
let us call this the value of that action.

Simple-average Method

Simple-average Method
𝑄𝑡(𝑎) converges to 𝑞∗(𝑎)

Action-value Methods

- Greedy Action Selection Method
- 𝜀-greedy Action Selection Method
- Upper-Confidence-Bound(UCB) Action Selection Method

Greedy Action Selection Method
𝑎𝑟𝑔𝑚𝑎𝑥 𝑎 𝑓(𝑎) - a value of 𝑎 at which 𝑓(𝑎) takes its maximal value

Greedy action selection always exploits
current knowledge to maximize immediate reward

Greedy Action Selection Method’s disadvantage
Is it a good idea to select greedy action,
exploit that action selection
and maximize the current immediate reward?

𝜀-greedy Action Selection Method
Exploitation is the right thing to do to maximize
the expected reward on the one step,
but Exploration may produce the greater total reward in the long run.

𝜀 – probability of taking a random action in an 𝜀-greedy policy

Exploitation
Exploration

Upper-Confidence-Bound(UCB) Action Selection Method
ln 𝑡 - natural logarithm of 𝑡
𝑁𝑡(𝑎) – the number of times that action 𝑎 has been selected prior to time 𝑡
𝑐 – the number 𝑐 > 0 controls the degree of exploration

The idea of this UCB action selection is that
The square-root term is a measure of
the uncertainty(or potential) in the estimate of 𝑎’s value
The probability that the slot machine
may be the optimal slot machine

UCB Action Selection Method’s disadvantage
UCB is more difficult than 𝜀-greedy to extend beyond bandits
to the more general reinforcement learning settings
One difficulty is in dealing with nonstationary problems
Another difficulty is dealing with large state spaces

4. A simple Bandit Algorithm
- Incremental Implementation
- Tracking a Nonstationary Problem

Incremental Implementation
𝑄 𝑛 denote the estimate of 𝑅 𝑛−1‘s action value
after 𝑅 𝑛−1 has been selected n − 1 times

𝑄 𝑛+1 = 𝑄 𝑛 +
1
𝑛
𝑅 𝑛 − 𝑄 𝑛
holds even for 𝑛 = 1,
obtaining 𝑄2 = 𝑅1 for arbitrary 𝑄1

Unstable(↔Constant)
Available on stationary problem

The expression 𝑇𝑎𝑟𝑔𝑒𝑡 − 𝑂𝑙𝑑𝐸𝑠𝑡𝑖𝑚𝑎𝑡𝑒 is an error in the estimate.
The target is presumed to indicate a desirable direction
in which to move, though it may be noisy.

Tracking a Nonstationary Problem

Why do you think it should be changed from
1
𝑛
to 𝛼?

Why do you think it should be changed from
1
𝑛
to 𝛼?
We often encounter RL problems that are effectively nonstationary.
In such cases it makes sense to give more weight
to recent rewards than to long-past rewards.
One of the most popular ways of doing this is to use
a constant step-size parameter.
The step-size parameter 𝛼 ∈ (0,1] is constant.

?
= 𝑄 𝑛 + 𝛼𝑅 𝑛 − 𝛼𝑄 𝑛
= 𝛼𝑅 𝑛 + (1 − 𝛼)𝑄 𝑛
𝑄 𝑛 = 𝛼𝑅 𝑛−1 + 1 − 𝛼 𝑄 𝑛−1 ?
∴ 𝑄 𝑛 = 𝛼𝑅 𝑛−1 + (1 − 𝛼)𝑄 𝑛−1

?

∴ 𝑄 𝑛−1 = 𝛼 𝑅 𝑛−2 + 1 − 𝛼 𝑅 𝑛−3 + ⋯ + 1 − 𝛼 𝑛−3 𝑅1 + 1 − 𝛼 𝑛−2 𝑄1
1 − 𝛼 2
𝑄 𝑛−1 = 1 − 𝛼 2
𝛼𝑅 𝑛−2 + 1 − 𝛼 3
𝛼𝑅 𝑛−3 + ⋯
+ 1 − 𝛼 𝑛−1
𝛼𝑅1 + 1 − 𝛼 𝑛
𝑄1 ?
= 1 − 𝛼 2{𝛼𝑅 𝑛−2 + 1 − 𝛼 𝛼𝑅 𝑛−3 + ⋯
+ 1 − 𝛼 𝑛−3 𝛼𝑅1 + 1 − 𝛼 𝑛−2 𝑄1}

Sequences of step-size parameters often converge
very slowly or need considerable tuning
in order to obtain a satisfactory convergence rate.
Thus, step-size parameters should be tuned effectively.

In addition to a simple bandit algorithm,
there is another way to use the gradient method as a bandit algorithm

We consider learning a numerical 𝑝𝑟𝑒𝑓𝑒𝑟𝑒𝑛𝑐𝑒
for each action 𝑎, which we denote 𝐻𝑡(𝑎).
The larger the preference, the more often that action is taken,
but the preference has no interpretation in terms of reward.
In other wards, just because the preference(𝐻𝑡(𝑎)) is large,
the reward is not unconditionally large.
However, if the reward is large, It can affect the preference(𝐻𝑡(𝑎))

The action probabilities are determined according to
a 𝑠𝑜𝑓𝑡 − 𝑚𝑎𝑥 𝑑𝑖𝑠𝑡𝑟𝑖𝑏𝑢𝑡𝑖𝑜𝑛 (i.e., Gibbs or Boltzmann distribution)

𝜋 𝑡(𝑎) – Probability of selecting action 𝑎 at time 𝑡
Initially all action preferences are the same
so that all actions have an equal probability of being selected.

There is a natural learning algorithm for this setting
based on the idea of stochastic gradient ascent.
On each step, after selecting action 𝐴 𝑡 and receiving the reward 𝑅𝑡,
the action preferences are updated.
Selected action 𝐴 𝑡
Non-selected actions

1) What does 𝑅𝑡 mean?
2) What does (𝑅𝑡 − 𝑅𝑡)(1 − 𝜋 𝑡 𝐴 𝑡 ) mean?

1) What does 𝑅𝑡 mean?
2) What does (𝑅𝑡 − 𝑅𝑡)(1 − 𝜋 𝑡 𝐴 𝑡 ) mean?
?

What does 𝑅𝑡 mean?
𝑅𝑡 ∈ R is the average of all the rewards.
The 𝑅𝑡 term serves as a baseline.
If the reward is higher than the baseline,
then the probability of taking 𝐴 𝑡 in the future is increased,
and if the reward is below baseline, then probability is decreased.
The non-selected actions move in the opposite direction.

What does (𝑅𝑡 − 𝑅𝑡)(1 − 𝜋 𝑡 𝐴 𝑡 ) mean?
- Stochastic approximation to gradient ascent
in Bandit Gradient Algorithm
- Expected reward
- Expected reward by Low of total expectationE[𝑅𝑡] = E[ E 𝑅𝑡 𝐴 𝑡 ]

?

?
The gradient sums to zero over all the actions, σ 𝑥
𝜕𝜋 𝑡(𝑥)
𝜕𝐻𝑡(𝑎)
= 0
– as 𝐻𝑡(𝑎) is changed, some actions’ probabilities go up
and some go down, but the sum of the changes must be
zero because the sum of the probabilities is always one.

?
E[𝑅𝑡] = E[ E 𝑅𝑡 𝐴 𝑡 ]

Please refer page 40 in link of reference slide

∴
We can substitute a sample of the expectation above
for the performance gradient in 𝐻𝑡+1 𝑎 ≅ 𝐻𝑡 𝑎 + 𝛼
𝜕E[𝑅 𝑡]
𝜕𝐻𝑡(𝑎)

6. Summary
In this chapter, ‘Exploitation & Exploration‘ is the core idea.

6. Summary
- Greedy Action Selection Method
- 𝜀-greedy Action Selection Method
- Upper-Confidence-Bound(UCB) Action Selection Method

6. Summary
A simple bandit algorithm : Incremental Implementation

6. Summary
A simple bandit algorithm : Tracking a Nonstationary Problem

6. Summary
Gradient Bandit Algorithm

6. Summary
A parameter study of the various bandit algorithms

Reinforcement Learning is LOVE♥

Multi-armed Bandits

More Related Content

What's hot (20)

Similar to Multi-armed Bandits (20)

More from Dongmin Lee (14)

Recently uploaded (20)

Multi-armed Bandits