-
1.
Reinforcement
Learning
The Exploding Kittens
Edition
Tarek Amr
-
2.
Why Reinforcement Learning?
I learned after
playing many times;
That I‘m more likely to
win if I played this move
after that one.
No one kept telling me
make this or that move!
-
3.
States, Actions and Rewards
St St+1
At At+1 St+2
Goal State
R
-
4.
What’s a good reward
If getting an
Exploding Kitten card
gives me a reward of
-1;
What reward do I get
if I get a Defuse card?
And for a Nope card?
-
5.
From Rewards, States get Values
And from
values comes
policies!
-
6.
a State has a value (V)
St St+1
At At+1 St+2
Goal State
R
Vt Vt+1
-
7.
or State/Action pair have a value (Q)
St St+1
At At+1 St+2
Goal State
R
Qt Qt+1
-
8.
Temporal Difference; S-A-R-S-A
St St+1
At At+1 St+2
Goal State
R
Qt := Qt + α (Rt+1 + γ Qt+1 - Qt)
-
9.
Epsilon Greedy
St
St+1At At+1 St+2
Goal State
RExploration vs Exploitation
Qt := Qt + α (Rt+1 + γ Qt+1 - Qt)
-
10.
Deep Q Learning
State Feature1 State Feature2 Action Value
10 20 JUMP 0.5
20 15 DUCK 0.6
15 25 JUMP 0.8
Warning:Over simplification Ahead
This is a Q-Table;
What if there are too many States & Actions?
-
11.
MDP, MC and TD
Markov Decision Process:
● You need to know the states and the transitions between them.
Monte Carlo (variance ↑):
● You wait till episode’s end, and re-assign values to states.
● No need to even know the states, we sample from the environment.
Temporal Difference (bias ↑):
● Update on the go. No need to even have goal states.
-
12.
Let’s play the RL vs SL game
for (i=0; i<3; i++) {
● Pick a catawiki problem
● Should it be solved via
○ Reinforcement learning?
○ Supervised learning?
}
We expect, in general, that the environment will be nondeterministic; that is, that taking the same action in the same state on two different occasions may result in different next states and/or different reinforcement values. However, we assume the environment is stationary; that is, that the probabilities of making state transitions or receiving specific reinforcement signals do not change over time.
Reinforcement learning differs from the more widely studied problem of supervised learning in several ways. The most important difference is that there is no presentation of input/output pairs. Instead, after choosing an action the agent is told the immediate reward and the subsequent state, but is not told which action would have been in its best long-term interests. It is necessary for the agent to gather useful experience about the possible system states, actions, transitions and rewards actively to act optimally.
Another difference from supervised learning is that on-line performance is important: the evaluation of the system is often concurrent with learning.
Use cases for RL: if there is path dependence (i.e. the order of your moves matter, like in chess), if you have a budget (e.g. max # emails to send, money), or if your decisions select your future training examples (e.g. (greedily) not bidding on new websites in programmatic advertising will never allow you acquire data about them). (via Peter Tegelaar)