Lecture 18
Video
- hill climbing, random restart hill climbing, iterative hill climbing
- search moves are deterministic
- steepest wle ki taraf hi jaunga
Random Walk
- randomly picks a neighbor and move towards it
- hill climbing with same starting point will always take me to same point
- but random walk might take at diff places in diff iterations
- these 2 are extremes
Exploitation and Exploration
- hill climbing me I am doing is exploitation
-
I am only relying on gradient
- Random walk me no concept of gradient
- I am doing exploration there
Simulated Annealing
- whwen temp is very high, molecules move very fast, when very down, they move slowly
- start with high temp, then follow a cooling schedule
- done by a probabilistic methid
- c is current state , n is next state, delta e is energy part = gain
- sigmoidal function (value b/w 0 and 1), S shaped
- t -> infinity, P(c,N) = 1/2
- if t-> 0 value = 1
- if P = 1, there is no randomness, puri certainity hai na bhai
- P = 0, again no randomness
- when I move towards 1/2 from both side, randomness increases
- so t->infinity = total randomness = random walk = total exploration
- t->0 = explotation = no randomness
- in sa, I start with exploration, slowly introduce exploitation, or vice versa?
- random restart hill climbing achi nahi bcz basis of attraction (range of points near global max) is bahut chotu sa
- yaha simulated annealing hi best hai