search problem

Search problems

consist

1. a state space: all possilbel states in the world
   world state: all info about the state in the world
   search state: only info needed for planning
   e.g. Pacman pathing / pacman eating dots
2. successor/result function:
   input: action,state
   output: successor state,cost
3. start state： agent exist initially
4. goal test: whether ends

一些概念

plan: a path from the start state to the goal state(search problem 的目标)
strategy: state 的考虑/搜索顺序 e.g. DFS,BFS

state space size

fundamental enumerate principle: variable domains‘ size multiply together
如果我们有n个变量，每个变量的domain大小分别为$x_1,x_2,x_3,\cdots$
那么状态空间的大小$x_1\ast x_2\ast \cdots$

state space graphs and search trees（搜索问题的图表示）

state space graphs

1. node: states
2. directed edge: actions,cost
   weight: the cost of actions(和边相关)
   each state only occurs once

search trees:

node show plans from the root to achieve the node
root node: start state
children: successors
states cans occur many times
search trees are greater or equal size to the state space graphs
很多时候（循环），我们没有办法建立整个搜索树
successor function: only save the states we are working with, and use the successor function to compute new ones（节省存储空间）

注： goal test 很多时候就是一个terminal nodes的集合

搜索图和搜索树都太大而不能全部存储

uninformed search

fringe: the observed but unvisited node
expand: pop a node from the fringe
generate: get neighbors of the popped node
stategy： 先pop什么节点

search algorithm properties:

1. completeness: guarantee to find a solution
2. optimal: find the lowest cost path
3. time complexity: generate nodes
4. space complexity:store nodes（size of fringe）
   branching factor b , 树的最大深度 m , 最优solution的深度 s
   $o(b^m)$nodes in a tree

depth first search

strategy: expand the deepest node first
implementation: fringe is LIFO stack
property:

1. completeness:❌ m can be infinite(循环) 永远找不到解
2. optimal:❌ find the depth path
3. time complexity: exploring the whole tree $o(b^m)$才能找到解
4. space complexity:$o(bm)$每层只扩展b个因子

breadth first search

strategy: expand the shallowest node first
implementation: fringe is FIFO stack(queue)
property:

1. completeness:✅
2. optimal: ✅前提不考虑每次拓展的cost，认为每个edge cost一致
3. time complexity: exploring the whole tree $o(b^s)$
4. space complexity:$o(b^s)$

expand 结束 do the goal test:
time complexity:$o(b^{s+1})$
space complexity:$o(b^{s+1})$

bfs robust but need more storage spaces

iterative deepening:
run a dfs with a depth limited 1,if no solution run a dfs with a depth limited 2
expand more nodes than bfs:
$b,b+b^2,b+b^2+b^3=bd+(d-1)b^2+\cdots =o(b^d)$
save storage space:$o(bd)$

uniform cost search

strategy: expand the lowest cost node first
implementation: fringe is priority queue (heap)
weight : path cost from the start node to the node or backward cost of v
从根节点扩展到该节点的开销
property:

1. completeness:✅
2. optimal: ✅（前提edge cost 不为负）如果有环就会导致来回走
   assume edge costs not negative
3. time complexity: optimal path costs:$C^{\ast }$,least arc cost$ϵ$ $o(b^{C^{\ast }/ϵ})$
   expand 所有比cheapest goal浅的节点
4. space complexity:$o(b^s)$ $o(b^{C^{\ast }/ϵ})$

sum up

uninformed search	completeness	optimal	time complexity	space complexity
bfs	✅	✅	$o(b^s)$	$o(b^s)$
dfs	❌	❌	$o(b^m)$	$o(bm)$
ucs	✅	✅	$o(b^{C^{\ast }/ϵ})$	$o(b^{C^{\ast }/ϵ})$

informed search

我们可以知道一点离目标state多近的信息

Heuristics Function

估计当前state与目标state的距离cost
Heuristics Function
input: state
output: the estimate of the distance between state to the goal state
usually want a lower bound
一般要比真实的少一点
用一个relaxed problem去估计启发式函数

greedy search

description: expand the node with lowest heuristics value
fringe: prioritiy queue (estimate forward value)
property:

1. completeness:not guaranteed
2. optimal: not guaranteed
   主要取决于启发式函数的选择，有可能直接找到路径，也有可能永远找不到

A* search

description: expand the node with lowest total estimated cost
fringe: prioritiy queue (estimate total value)
$f(n)=g(n)+h(n)$
property:
3. completeness:✅
4. optimal: ✅
前提，启发式函数要满足一些性质
有点：速度很快，然后可以保证完备和最优
坏的启发式会破坏：$h(n)=1-g(n)$$f(n)=1$
对于边cost不一致的情况，会被破坏掉

Admissiblity and consistency

admissible:$0\le h(n)\le h^{\ast }(n)$
$ h^*(n)$到达goal state的最小值
theorem: for a given search problem, if the admissibility constraint is satisfied by a heuristic function h, using $A^{\ast }$ tree search will yield an optimal solution.
proof:
A is the optimal goal state, B is a suboptimal goal state;n(include A) are ansesters of A, we claim that n must be chosen before B

1. $g(A)<g(B)$
2. $h(A)=h(B)=0$
3. $f(n)\le f(A)$ $f(n)=h(n)+g(n)\le h^{\ast }(n)+g(n)=g(A)=f(A)$
   $f(A)<f(B)\to f(n)<f(B)$

how to design admissible heuristic function

1. from a relaxed problem
2. from a subproblem

一些不一致的启发式函数也有可能很好用，因为可能减少了计算时间
graph search:
maintain a closed set for expanded nodes to avoid expand twice

closed must be a set, since checking list takes $o(n)$ time
completeness: if exists , will be visited
optimality:however, graph search may ruin the optimality of $A\ast$, even the heuristic function is admissibility
e.g.

consistency:$\forall A,C$ $h(A)-h(C)\le cost(A,C)$
theorem: for a given search problem, if the consistency constraint is satisfied by a heuristic function h, using $A^{\ast }$ graph search will yield an optimal solution.

proof:

$h(n)\le h(n^{\prime })+cost(n,n^{\prime })$
$n^{\prime }$ is a successor of n
$f(n^{\prime })=h(n^{\prime })+g(n^{\prime })=g(n)+h(n)+cost(n,n^{\prime })\ge g(n)+h(n)=f(n)$
so f(n) is nondecreasing along the path
so whenever a node is removed from the fringe, it’s optimal path has already been found
so we only need to show, optimal solution A will be found before B
$h(A)=h(B)=0$
$f(A)<f(B)$

Under admissibility and consistency, $h(A)=h(B)=0$
consistency implies adimissiblity
图搜索能保证 ucs的最优性和完备性，因为ucs $h=0$既一致又最优

diminance

dominant：$\forall n{h}_a(n)\ge h_b(n)$
trivial heuristic: h(n)=0，reduce $A^{\ast }$ to $UCS$
max of admissible heuristics is admissible
可以写多个一致或可达的启发式函数取max，仍然可以保证一致性可可达性

$A^{\ast }$的computation cost会多一些
bfs不会有cycle问题，因为我们会访问比他们的相邻节点更早，cycle会有更大的cost
图搜索不会破坏完备性，但却会破坏最优性

图搜索对于dfs
可以破掉循环，可以保证完备性
图搜索对于bfs
可以保证最优性和完备性

题目复习

$A^{\ast }$可以保证比 ucs 产生更少的nodes
注意搜索问题性质改变会导致是否能够解改变，搜索图没有办法完全决定能不能求解