ECS 170:
Introduction to Artificial Intelligence
Prof. Rao Vemuri
Central
Issues in AI (This is NOT a course outline)
1. Methods of Problem Solving
· Generate and Test Methods
· Divide and Conquer Methods
· Systematic Methods
2. Searching
for a Solution.
State Space
Search. Search Methods dominate AI. Intuitively, a state is represented by a
set of related variables along with their values. Blindly searching through
state space is hard. Often the search is assisted by rules of thumb or
"heuristics". Heuristic search, although better than blind search, is
also fallible.
Search Methods
Blind Search Methods
· Depth-first Search (DFS)
· Breadth-First Search (BFS)
· Random Search
Heuristic Search Methods
Hill Climbing = DFS + Measurements
· Steepest Ascent Hill Climbing (SAHC)
· Next Ascent Hill Climbing (NAHC)
· Random Mutation Hill Climbing
· A* Search
Special Search Methods
· Simulated Annealing
· Genetic Algorithms for Optimization
· Genetic Programming for Planning
· Beam Search - Expands Several Partial Paths and Purges the rest
· Best First Search
· Branch and Bound Search, expands least cost partial path
· A* Search, is a combination of B&B and Dynamic Programming ideas. The heuristic makes estimates of the remaining distance.
Adversarial Search and Game Trees
· Minimax search
· Alpha-beta pruning
· Static Evaluation Function
· Horizon effect
· Progressive deepening until the allocated time is expended
3. Knowledge Representation
There are many ways of representing knowledge.
· Semantic nets
· Propositional Logic
· First Order (Predicate)
Topics in Logic include syntax, semantics, models, logical
entailment, proofs, soundness, completeness, and decidability. Topics in Reasoning Methods include the truth table method,
natural deduction, the
Davis-Putnam procedure,
resolution, model elimination, demodulation, and paramodulation. We will not cover all of this stuff.
Logic is one of the most fundamental knowledge representation schemes we know. A predicate can also be represented by a symbol. This symbol has a truth value. A predicate of “arity” n can take on n ordered arguments or terms. For example, the statement "The fourth quarter profit in HP is high" can be represented by the predicate "profit" whose arguments are hp, 4q, and high. This is compactly written as
profit (hp, 4q, high)
We can make the arguments into variables and write -
profit (Company-name, Quarter, Performance)
4.
Knowledge-based Systems:
If we have plenty of knowledge to start with, we may avoid searching and find a
solution using knowledge-based approach. An expert
system is a special case of a knowledge-based system in which domain-specific
expert knowledge is available. If no such domain specific expertise is
available, then it is a weak method. Many knowledge-based systems use
the so-called production model, which uses production rules. The
long-term memory in which the production rules are stored is called the knowledge
base.
5. Machine Learning
Knowledge is usually not static. So the knowledge base can easily get outdated.
So learning is necessary to adapt to changing surroundings. Learning refers to
positive changes toward improved performance. In symbol-based machine
learning, one searches through concept space to find the desired
concept. As opposed to symbol-based methods, those using artificial neural nets
are called sub-symbolic methods.
· Neural Nets
· Bayesian Nets
Department of Computer Science
University of California at Davis
Davis, CA 95616-8562