Artificial Intelligence

Beyond classical search

B E (CSE) B Section, Semester 5

T. T. Mirnalinee
Department of Computer Science and Engineering
SSN College of Engineering
24th August 2020
Introduction
• Observable
• Deterministic
• Known environments
• Local search: rather than systematically exploring- explore/evaluate
from current state
• Simulated annealing
• Genetic Algorithm
Local Search
• Local search -- one or few current nodes in memory but exhaustive search
maintains all the successors
• Takes very less memory – very good in very large or infinite state spaces
• State space is continuous in several ML problems
• It may find reasonable solution (with respect to optimization)
• It is used both in search and optimization problem
• Search a state which has the max/min objective function value
• Profit – Max, error/cost – Min. Maximization can be converted to
minimization
Local search
• Informed and uninformed search – search all states – stores all the
states from initial states to the solution
• Local search – path to the goal is not relevant
• Eg – 8 queens problem only the final configuration is sufficient
• These different class of algorithms are local search – it focus on
current state and neighbouring state
• Paths are not retained
• Less memory, state space is large – it could able to find reasonable
solution
Local search
• Local search can be used if we have heuristic function – so informed search
• Hill climbing
• Simulated annealing
• Genetic algorithm
• Local maxima - process may end even better solution exist - backtracking
• Plateau – all neighbours have same value, - big jump – randomly select a
state far away from the current state
• Ridges – several peaks in the neighbour - move simultaneously in various
direction and select the best
State space landscape

Shoulder – improvement is possible

Plateau – No improvement

• Local search algorithm – explore the

landscape to find the global
maximum/minimum
Local Search Algorithms and optimization
problems
• Optimization – objective function
• Local search – state is a solution
• No of pairs of queens that are attacking each other (objective function)
• Solution – expected ( 0 – so minimization problem)
• nxn board – we do have nn solution (state)
• Local search we are searching in a solution space
• States are independent of each other
• Solution to systematic search: path cost till the goal, order in which states
are arranged(8-queens)
• Local search: find goals as wells as suitable for optimization problems-
finding the best state subject to an objective function.
• Complete and optimal
8-queens problem
• Complete state formulation – each state has eight queens on board,
one per column
• Successor – moving a single queen to another square in the same
column
• A state has 56 sucessors (8x7)
• Heuristic function number of pairs of queens that are attaching each
other – minimization problem
• Hill climbing algorithm choose randomly among the set of successors
if there is more than one
Random Restart
• Trivial algorithm – random sampling a new state
• Take one queen put in first column, another queen second column
and so on
• If we keep doing this at some point we will get an optimal solution
• Random walk: start from a state, randomly jump to any one of the
neighbour
• At some point we will get a solution provided only one connected
component
• Every state has a neighbour, greedy search – out of many alternatives
choose the best and moves
Hill climbing
• Greedy local search – search the good neighbour state
• Local maxima: peak higher than the neighbouring states but lower
than the global maximum
• Ridges: Ridges result in a sequence of local maxima that is very
difficult for the greedy to navigate
• Plateau : flat area of the state space landscape
• Trapped in local minima – so it is incomplete
Hill-climbing search(Greedy local search)
• Steepest ascent
• Continuously moves in the direction of increasing value (uphill)
• Terminates when it reaches the peak – no neighbour with higher value
• Greedy local search – grab the good neighbour without bothering about where to go
next
function HILL-CLIMBING(problem) returns a state that is a local
maximum
current ←MAKE-NODE(problem.INITIAL-STATE)
loop do
neighbor ←a highest-valued successor of current
if neighbor.VALUE ≤ current.VALUE then return current.STATE
current←neighbor
 Variant of hill climbing -- incomplete
• Hill climbing is incomplete – that is algorithm will lead to a neighbour
with better current state
• Greedy HC – always select a best neighbour
• Stochastic – chooses at random from the uphill moves
• Select any better neighbour randomly with a probability distribution
• Every Current state has large neighbours memory problem still exist
• First-choice – randomly generate successor until the one which is
better than the current state (Select only a subset of successor)
• Random-restart (complete) – series of search until a goal is found, if
few local maxima and plateau it will find the solution quickly.
Variant of hill climbing
• Randomly generate a state and do hill climbing
• Again generate a state and do hill climbing
• Repeat it
• If randomly selected state is goal – it is complete
Nature inspired algorithm
• Designing algorithms from nature – nature inspired/biological
algorithm
• Model the problem as the landscape
• Local minimum vs global minimum
• Many local minimum but only one global minimum
Simulated Annealing
• Combining hill climbing with random walk – achieve efficiency and
completeness
• Annealing is the process used to temper or harden metals/glass by heating
them to a high temperature and gradually cooling them
• Reach low energy crystalline state
• This is achieved by gradient descent- shake the surface if it get stuck in
local minimum,
• imagine the task of getting a ping-pong ball into the deepest valley in a
bumpy surface
• Shake in such a way that it has to come out from local minimum but not
from global minimum
Simulated Annealing
• Intensity of shaking is temperature
• Shaking hard (high temperature)
• Decrease the intensity of shaking(gradually decrease the temperature)
• Instead of picking the best neighbour, here look for a random
neighbour
• Random move is picking a good solution accept or accept with a
probability of less than one
• Complete and efficient
 Simulated annealing- Algorithm
function SIMULATED-ANNEALING(problem, schedule) returns a
solution state
inputs: problem, a problem schedule , a mapping from time to
“temperature”
current ←MAKE-NODE(problem.INITIAL-STATE) P decreases with decreas
for t = 1 to∞do inΔE,
T ←schedule(t ) P decreases with decrease
At high values of T bad mov
if T = 0 then return current
are not accepted
next←a randomly selected successor of current
As T decreases, bad moves
ΔE ←next.VALUE – current.VALUE less likely to be accepted
if ΔE > 0 then current ←next Global minimum with
else current ←next only with probability eΔE/T probability of 1
Find global optimum with probability approaching 1
Simulated Annealing
• T= 0, no need to shake anymore
• Randomly generate successor (not all successor)
• If the next state is best value move otherwise discard and select a
new successor
• Here in SA, even if the successor is worse, no need to discard all the
time, sometimes we will take the successor
Local beam search

• Keeps tracks of k states rather than one (if memory is not the
limitation)
• Starts with k randomly generated states
• Each step all the successors of all the k states are generated.
• If any one is goal algorithm stops otherwise select the k best
successor
• Random restart search – each search process runs independently
• Local beam search- useful information is passed among the
parallel search
Local beam search
• Resemble random restart (randomly generate one initial state rum HC
and repeat)
• But in Local beam search – generate k states initially (appears to be
parallel)
• All K current state share the useful information among all the k states
(select the best K from all the successors)
Stochastic beam search

• Select any k states random

• Probability of selecting a successor is proportional to its fitness value
• Genetic Algorithm
• Resembles natural selection
• Organism – states
• Population – set of states
• Offsprings – successors
• Values – fitness
• Select next state based on the fitness
Genetic Algorithm
• Hard optimization problem can be solved using nature inspired
algorithm
• How biological process evolved
• Two states are combined to generate a new successor
Genetic algorithm
• New state is generated by combing two parent state
• GAs begin with a set of k randomly generated states, called the
population.
Each state, or individual, is represented as a string over a finite alphabet
Every state is represented by a string of symbols from alphabet (generally bit
string/digit string 0’s and 1’s)
8 –queens problem
• Instance - 24613578
GA
• Fitness function
• Each state is evaluated by the fitness function
• 8- queens pbm : no of non attacking queens
GA Operation
• Selection
• Crossover
• Mutuation
• 24, 23, 20,11 are fitness value (no of non attacking pairs)
• S= 24+23+20+11
• P= 24/S….
• Out of the four individuals are selected , one can be selected more
than once as well
• Cross over
• Crossover points are selected randomly
• Initially can take larger points
• Later when near to the global minima, smaller points
Genetic algorithm
function GENETIC-ALGORITHM(population, FITNESS-FN) returns an individual
inputs: population, a set of individuals FITNESS-FN, a function that measures the fitness of an
individual
repeat
new population ←empty set
for i = 1 to SIZE(population) do
x ←RANDOM-SELECTION(population, FITNESS-FN)
y ←RANDOM-SELECTION(population, FITNESS-FN)
child ←REPRODUCE(x , y)
if (small random probability) then child ←MUTATE(child )
add child to new population
population ←new population
until some individual is fit enough, or enough time has elapsed
return the best individual in population, according to FITNESS-FN
function REPRODUCE(x , y) returns an individual
inputs: x , y, parent individuals
n←LENGTH(x ); c←random number from 1 to n
return APPEND(SUBSTRING(x, 1, c), SUBSTRING(y, c + 1, n))
Genetic algorithm
• Combines better neighbour and random exploration
• Sharing of information from all the states in the population
• Granularity of the search
• schema, which is a substring in which some of the positions can be
left unspecified -246*****
• Since 246 is a safe substring * can be replaced by any value
• Schema with best fitness value will sustain
Summary
• For hard optimization problem
• Circuit layout
• Job scheduling

• ACO – how to effectively find food

• PSO – birds find path