Problem-Solving as Search
Intelligent Agents
Agent: Anything that can be viewed as perceiving its environment through sensors and acting upon that environment through actuators. Agent Function: Agent behavior is determined by the agent function that maps any given percept sequence to an action. Agent Program: The agent function for an artificial agent will be implemented by an agent program.
A Simple Reflex Agent
Agent Sensors
What the world is like now
Environment
Condition-action rules
What action I should do now
Actuators
Agent with Model and Internal State
Agent Sensors
How the world evolves
What the world is like now
Environment
Condition-action rules
What action I should do now
Actuators
Goal-Based Agent
Agent Sensors What the world is like now What it will be like if I do action A
Environment
How the world evolves
Goals
What action I should do now Actuators
Schedule
Search Machine learning Knowledge based systems Discovery
Problem Solving as Search
Search is a central topic in AI
Originated with Newell and Simon's work on problem solving. Famous book: Human Problem Solving (1972)
Automated reasoning is a natural search task More recently: Smarter algorithms
Given that almost all AI formalisms (planning, learning, etc.) are NP-complete or worse, some form of search is generally unavoidable (no smarter algorithm available).
Defining a Search Problem
State space - described by initial state - starting state actions - possible actions available successor function; operators - given a particular state x, returns a set of < action, successor > pairs Goal test - determines whether a given state is a goal state (sometimes list, sometimes condition). Path cost - function that assigns a cost to a path
The 8 Puzzle
3 Initial State
6 Goal State
Clicker
What is the size of the state space?
A. 4 B. 3x3 C. 9! D. 99 E. Whatever
Clicker
How many actions possible for each state (on average)?
A. ~1 B. ~4 C. ~9 D. ~9!
Cryptarithmetic
SEND + MORE -----MONEY Find (non-duplicate) substitution of digits for letters such that the resulting sum is arithmetically correct. Each letter must stand for a different digit.
Solving a Search Problem: State Space Search Input:
Initial state Goal test Successor function Path cost function
Output:
Path from initial state to goal state. Solution quality is measured by the path cost.
Generic Search Algorithm
L = make-list(initial-state) loop node = remove-front(L) (node contains path of how the algorithm got there) if goal-test(node) == true then return(path to node) S = successors (node) insert (S,L) until L is empty return failure
Search procedure defines a search tree
Search tree root node - initial state children of a node - successor states fringe of tree - L: states not yet expanded
Search strategy - algorithm for deciding which leaf node to expand next. stack: Depth-First Search (DFS). queue: Breadth-First Search (BFS).
Solving the 8-Puzzle
5 4 1 2 3 6 1 8 8 4
3 Start State
6 Goal State
What would the search tree look like after the start state was expanded?
Node Data Structure
PARENT-NODE NODE
STATE
ACTION= right DEPTH=6 PATH-COST=6
2 CHILD-NODE CHILD-NODE
Sliding Block Puzzles
8-puzzle (on 3x3 grid) has 181,440 states
Easily solvable from any random position
15-puzzle (on 4x4 grid) has ~1.3 Trillion states
Solvable in a few milliseconds
24-puzzle (on 5x5 grid) has ~1025 states
Difficult to solve
Evaluating a Search Strategy
Completeness: Is the strategy guaranteed to find a solution when there is one? Time Complexity: How long does it take to find a solution? Space Complexity: How much memory does it need? Optimality: Does strategy always find a lowest-cost path to solution? (this may include different cost of one solution vs. another).
Uninformed search: BFS
Consider paths of length 1, then of length 2, then of length 3, then of length 4,....
Time and Memory Requirements for BFS O(bd+1)
Let b = branching factor, d = solution depth, then the maximum number of nodes generated is: b + b2 + ... + bd + (bd+1-b)
Time and Memory Requirements for BFS O(bd+1)
Example: b = 10 10,000 nodes/second each node requires 1000 bytes of storage
Depth Nodes 2 4 6 8 10 12 14 1100 111,100 107 109 1011 1013 1015
Time .11 sec 11 sec 19 min 31 hrs 129 days 35 yrs 3523 yrs
Memory 1 meg 106 meg 10 gig 1 tera 101 tera 10 peta 1 exa
Uniform-cost Search
Use BFS, but always expand the lowest-cost node on the fringe as measured by path cost g(n). s s
0
A A 1 s 15 C 5 B 5 5 10 A G B 11 A B G C 15 5 C 15 s 1 B 5 C 15 s
Requirement: g(Successor(n))
g(n)
11
10
Always expand lowest cost node in open-list. Goal-test only before expansion, not after generation.
Uninformed search: DFS
DFS vs. BFS
Complete Optimal Time Space
BFS
DFS
YES
Finite depth
YES
NO
O(bd+1)
O(bm)
O(bd+1)
O(bm)
m is maximum search depth d is solution depth b is branching factor
Time m = d: DFS typically wins m > d: BFS might win m is infinite: BFS probably will do better Space DFS almost always beats BFS
Which search should I use...
If there may be infinite paths?
B=BFS
D=DFS
Which search should I use...
If goal is at a known depth?
B=BFS
D=DFS
Which search should I use...
If there is a large (possibly infinite) branching factor?
B=BFS
D=DFS
Which search should I use...
If there are lots of solutions?
B=BFS
D=DFS
Backtracking Search
Idea: DFS, but dont expand all b states before next level Generate the next state as needed (e.g. from previous state) Uses only O(m) storage Important when space required to store each state is very large (e.g. assembly planning)
Iterative Deepening [Korf 1985]
Idea: Use an artificial depth cutoff, c. If search to depth c succeeds, we're done. If not, increase c by 1 and start over.
Each iteration searches using depth-limited DFS.
Limit=0
Iterative Deepening
Limit=1
Limit=2
Limit=3
Cost of Iterative Deepening
space: O(bd) as in DFS, time: O(bd)
b 2 3 5 10
ratio of IDS to DFS 3 2 1.5 1.2
25
100
1.08
1.02
Bidirectional Search
Comparing Search Strategies
Criterion Breadth -First bd+1 bd+1 Yes Yes UniformCost
b b
1 C* C*
DepthFirst bm bm no No
Iterative Deepening bd bd yes Yes
Bidirectional (if applicable) bd/2 bd/2 yes Yes
Time Space Optimal? Complete?
yes Yes
***Note that many of the ``yes's'' above have caveats, which we discussed when covering each of the algorithms.
