State Space Search
1. The Problem
AI as problem-solving, and problem-solving as state-space-search.
blind (uninformed, brute-force) search: DFS and BFS, forward and backward.
resource demand of search, "combinatorial (exponential) explosion".
2. Proposed Solutions
partial information and heuristic (best-first) search, A* algorithm
Newell-Simon76: search as the core of AI. General Problem Solver (GPS): means-ends analysis.
game playing by search: minimax algorithm and alpha-beta pruning
Deep Blue: heuristic search with parallel hardware, opening book and endgame database, and complicated evaluation function.
3. Issues
(1) no efficient heuristic function
(2) problem-solving cannot be naturally represented as search, which is based on several assumptions
4. Readings
Chapter 3, Sections 4.1-2, 6.1-4
Newell-Simon76
5. Ideas
AI as "problem solving" as "heuristic search"
reasonable:
- cognitive processes are all goal-directed
- a solution usually consists of many basic operations/actions
- when looking for a solution, the effect of every action needs to be predicted
- choice among candidate solutions is based on quantitative evaluation and comparison
problematic:
- to represent domain knowledge as a state space (graph) [to allow incomplete description]
- to require a list of applicable actions [to use reasoning to reveal possibilities]
- to apply a predetermined heuristic function [to learn control knowledge and preference]
- to expect certain consequences [to expect certain consequences to various degrees]
- to assume that all state changes are caused by the agent [to keep observing]
- to follow one path at a time [to follow multiple paths at different speeds by distributing resources; parallel terraced scan]
- to stop at goal state [anytime algorithm; to stop when exhausted, with various levels of satisfaction]
common themes: facing uncertainty, unifying processes, managing resources