REINFORCEMENT LEARNING

REINFORCEMENT LEARNING
NAME : S. THARABAI
COURSE: M.TECH(CSE) II YEAR PT
REG. No : 121322201011
REVIEW No: 2

WHAT IS REINFORCEMENT
LEARNING?

REINFORCEMENT IS DEFINED AS
CHARACTERISING THE
LEARNING PROBLEM

REINFORCEMENT LEARNING
WORKS ON PRIOR AND
ACCURATE MEASUREMENTS
IT WORKS ON SOME
CALCULATIONS AND PLACES
WHERE HUMAN CAN'T REACH. IT
IS INTANGIBLE

IT IS NOT A REAL WORLD
PROBLEM BUT TRYING TO
REACH THINGS BEYOND THE
WORLD!

Reinforcement learning is
learning what to do--how to map situations to actions--so as to
maximize a numerical reward signal.
These two characteristics--trial-and-error search and delayed
reward--are the two most important distinguishing features of
reinforcement learning.

EXAMPLES
➢ CHESS BOARD MOVEMENTS
➢ ROBOTICS MOVEMENTS
➢ DEPTH OF THE PETROLEUM REFINERY
AND ADAPTOR

Comparison Between Machine Learning
And Reinforcement Learning.
Machine learning works classified and unclassified data sets.
Reinforcement learning requires prior – accurate measurements.
Example, Robotics Arms should be measured and given, how many
steps to move and in what direction based on temperature adaptors.
Machine learning is a real-world problem. Reinforcement learning
works on some calculations, and helps in places where man can not
reach. Example, Petroleum Refineries.
Machine learning is tangible. Reinforcement learning is intangible.
Example, Chandrayan-1 project.
Machine learning works on sequence of step, reinforcement learning
works on episodesepisodes

ELEMENTS OF REINFOREMENT
LEARNING
➢ A POLICY
➢ A REWARD FUNCTION
➢ A VALUE FUNCTION
➢ A MODEL ENVIRONMENT

POLICY, REWARD AND VALUE
➢ POLICY DEFINES LEARNING AGENTS'S
WAY OF BEHAVING AT A GIVEN TIME.
➢ REWARD DEFINES A GOAL IN A
REINFORECEMENT PROBLEM
➢ VALUE FUNCTION SPECIFIES WHAT IS
GOOD IN THE LONG RUN

ENVIRONMENT
The fourth and final element of some reinforcement learning systems is a
model of the environment. This is something that mimics the behavior of the
environment. For example, given a state and action, the model might predict
the resultant next state and next reward. Models are used for planning

An -Armed Bandit Problem
An analogy is that of a doctor choosing
between experimental treatments for a
series of seriously ill patients. Each play
is a treatment selection, and each
reward is the survival or well-being of
the patient.

In our -armed bandit problem,
Each action has an expected or mean reward given
that the action is selected; let us call this the value of
that action. If you knew the value of each action,
then it would be trivial to solve the -armed bandit
problem: you would always select the action with
highest value. We assume that you do not know the
action values with certainty, although you may have
estimates.

Monte Carlo Methods
Monte Carlo methods are ways of solving the reinforcement learning problem based on averaging
sample returns. To ensure that well-defined returns are available, we define Monte Carlo methods only
for episodic tasks. That is, we assume experience is divided into episodes, and that all episodes
eventually terminate no matter what actions are selected. It is only upon the completion of an episode
that value estimates and policies are changed. Monte Carlo methods are thus incremental in an
episode-by-episode sense, but not in a step-by-step sense. The term "Monte Carlo" is often used more
broadly for any estimation method whose operation involves a significant random component. Here we
use it specifically for methods based on averaging complete returns.

The backup diagram for Monte
Carlo estimation of

Approximate state-value functions for the
blackjack policy that sticks only on 20 or 21,
computed by Monte Carlo policy evaluation.

The spectrum ranging from the one-step backups
of simple TD methods(step-by-step value) to the
up-until-termination backups of Monte Carlo
methods(next episode reward)

Performance of -step TD methods. The performance measure
shown is the root mean-squared (RMS) error between the true
values of states and the values found by the learning methods,
averaged over the 19 states, the first 10 trials, and 100 different
sequences of walks.

A simple maze (inset) and the average learning curves
for Dyna-Q agents varying in their number of planning
steps per real step. The task is to travel from S to S as
quickly as possible.

Action-Value Methods
we denote the true (actual) value of action a as, Q*(a) and the
estimated value at the rth play as Qt(a) . Recall that the true value
of an action is the mean reward received when that action is
selected. One natural way to estimate this is by averaging the
rewards actually received when the action was selected. And r be
the number of plays in which Ka is the times prior to rth play.

yielding rewards, r1+r2+....rk
then its value is estimated to be,

DYNAMIC PROGRAMMING
The reinforcement learning is more useful in
tracking the Non-stationary problems.
In such cases it makes sense to weight recent
rewards more heavily than long-past ones. Ex.
DYNAMIC PROGRAMMING.

The Agent-Environment Interface
The reinforcement learning problem is meant to
be a straightforward framing of the problem of
learning from interaction to achieve a goal. The
learner and decision-maker is called the agent.
The thing it interacts with, comprising
everything outside the agent, is called the
environment.

The Markov Property
In the reinforcement learning framework, the
agent makes its decisions as a function of a signal
from the environment called the environment's
state. In particular, we formally define a property of
environments and their state signals that is of
particular interest, called the Markov property.

Example 1: Bioreactor Suppose reinforcement
learning is being applied to determine
moment-by-moment temperatures and stirring
rates for a bioreactor (a large vat of nutrients and
bacteria used to produce useful chemicals)

Example 2: Pick-and-Place Robot Consider using
reinforcement learning to control the motion of a
robot arm in a repetitive pick-and-place task. If we
want to learn movements that are fast and
smooth, the learning agent will have to control the
motors directly and have low-latency information
about the current positions and velocities of the
mechanical linkages.

Sample Remote RCXLisp Code Illustrating
Sensor-Based Motor Control
LISP CODING

(DEFUN full-speed-ahead (r s dir)
“This will make the rcx in stream R go at speed S
in direction DIR until touch sensor on its ‘2’ port returns 1.”
(LET ((result 0))
(set-effector-state ‘(:A :B :C) :power :off r)
;in case things are in an inconsistent state,
;turn everything off first
(set-effector-state ‘(:A :C) :speed s r)
(set-effector-state ‘(:A :C) :direction dir r)
; dir is eq to :forward, :backward, or :toggle
; no motion will occur until the
; next call to set-effector-state
(set-sensor-state 2 :type :touch :mode :boolean r)
(set-effector-state ‘(:A :C) :power :on r)
(LOOP ;this loop will repeat forever until sensor 2 returns a 1
(SETF result (sensor 2 r))
(WHEN (AND (NUMBERP result)
;needed to keep = from causing error if
;sensor function returns nil due to timeout.
(= result 1))
(RETURN)))
(set-effector-state ‘(:A :C) :power :float r))))
(DEFUN testing ()
(with-open-com-port (port :LEGO-USB-TOWER)
(with-open-rcx-stream (rcx10 port :timeout-interval 80 :rcx-unit 10)
; increase/decrease serial timeout value of 80 ms depending on
;environmental factors like ambient light.
(WHEN (alivep rcx10)
(full-speed-ahead rcx10 5 :forward)))))

Integrating Planning, Acting, and
Learning
When planning is done on-line, while
interacting with the environment, a number
of interesting issues arise. New information
gained from the interaction may change the
model and thereby interact with planning. It
may be desirable to customize the planning
process in some way to the states or decisions
currently under consideration, or expected in
the near future.

Relationshipsamonglearning,planning,andacting.

Dyna-Q, a simple architecture integrating the major functions
needed in an on-line planning agent. Each function appears in
Dyna-Q in a simple, almost trivial, form

The complete algorithm for Dyna-Q.

A simple maze (inset) and the average learning
curves for Dyna-Q agents varying in their number of
planning steps per real step. The task is to travel
from S to S as quickly as possible.

Policies found by planning and nonplanning Dyna-Q
agents halfway through the second episode. The arrows
indicate the greedy action in each state; no arrow is
shown for a state if all of its action values are equal. The
black square indicates the location of the agent.

Dimensions of Reinforcement Learning
The dimensions of the reinforcement learning have to do with
the kind of backup used to improve the value function. The
vertical dimension is whether they are sample backups (based
on a sample trajectory) or full backups (based on a distribution
of possible trajectories). Full backups of course require a
model, whereas sample backups can be done either with or
without a model (another dimension of variation).

A slice of the space of
reinforcement learning methods.

The horizontal dimension corresponds to the depth of backups,
that is, to the degree of bootstrapping. At three of the four
corners of the space are the three primary methods for
estimating values: DP, TD, and Monte Carlo.
A third important dimension is that of function approximation.
Function approximation can be viewed as an orthogonal
spectrum of possibilities ranging from tabular methods at one
extreme through state aggregation, a variety of linear methods,
and then a diverse set of nonlinear methods

In addition to the four dimensions just discussed, we have
identified a number of others throughout the book:
Definition of return
Is the task episodic or continuing, discounted or undiscounted?
Action values vs. state values vs. afterstate values
What kind of values should be estimated? If only state values
are estimated, then either a model or a separate policy (as in
actor-critic methods) is required for action selection.
Action selection/exploration
How are actions selected to ensure a suitable trade-off
between exploration and exploitation? We have considered
only the simplest ways to do this: -greedy and softmax action
selection, and optimistic initialization of values.

Synchronous vs. asynchronous
Are the backups for all states performed
simultaneously or one by one in some order?
Replacing vs. accumulating traces
If eligibility traces are used, which kind is most
appropriate?
Real vs. simulated
Should one backup real experience or simulated
experience? If both, how much of each?

Location of backups
What states or state-action pairs should be backed up?
Modelfree methods can choose only among the states
and state-action pairs actually encountered, but
model-based methods can choose arbitrarily. There are
many potent possibilities here.
Timing of backups
Should backups be done as part of selecting actions, or
only afterward?
Memory for backups
How long should backed-up values be retained? Should
they be retained permanently, or only while computing an
action selection, as in heuristic search?

Elevator Dispatching
Waiting for an elevator is a situation :
We press a button and then wait for an elevator to arrive traveling in the right
direction.
We may have to wait a long time if there are too many passengers or not enough
elevators.
Just how long we wait depends on the dispatching strategy the elevators use to
decide where to go.
For example, if passengers on several floors have requested pickups, which
should be served first? If there are no pickup requests, how should the elevators
distribute themselves to await the next request?
Elevator dispatching is a good example of a stochastic optimal control problem of
economic importance that is too large to solve by classical techniques such as
dynamic programming.

Four elevators in a ten-story building

Job-Shop Scheduling
Many jobs in industry and elsewhere require completing a collection of tasks
while satisfying temporal and resource constraints. Temporal constraints say that
some tasks have to be finished before others can be started; resource constraints
say that two tasks requiring the same resource cannot be done simultaneously
(e.g., the same machine cannot do two tasks at once). The objective is to create a
schedule specifying when each task is to begin and what resources it will use that
satisfies all the constraints while taking as little overall time as possible. This is
the job-shop scheduling problem.

JOB SHOP SCHDULING APPLICATION IN
NASA
The NASA space shuttle payload processing
problem (SSPP), which requires scheduling the
tasks required for installation and testing of
shuttle cargo bay payloads. An SSPP typically
requires scheduling for two to six shuttle
missions, each requiring between 34 and 164
tasks. An example of a task is
MISSION-SEQUENCE-TEST.

CONCLUSION
All of the reinforcement learning methods we have explored in this
ppt have three key ideas in common.
First, the objective of all of them is the estimation of value functions.
Second, all operate by backing up values along actual or possible
state trajectories.
Third, all follow the general strategy of generalized policy iteration
(GPI), meaning that they maintain an approximate value function
and an approximate policy, and they continually try to improve each
on the basis of the other.
We suggest that value functions, backups, and GPI are powerful
organizing principles potentially relevant to any model of
intelligence.

REINFORCEMENT LEARNING

More Related Content

What's hot (20)

Viewers also liked (12)

Similar to REINFORCEMENT LEARNING (20)

Recently uploaded (20)

REINFORCEMENT LEARNING