Brief 1: Artificial Intelligence Overview 01/20/95
The central ideas in Russell and Norvig's Part I are summarised below. Items which I consider especially useful for the exam are marked with an asterisk (*). Suggestions for discussion are included at the end.
Chapter 1: Introduction
1. The computer is a tool for testing theories of intelligence.
2 *. Table 1: Paradigms for AI Research
System Characteristics Think .. Act .. .. like humans Cognitive modeling: Turing test approach: requires experimentation relevant when computers on animals and humans. interact with people. most constrained approach .. rationally Logicist tradition: Agents achieve goals, formalize and operate on given beliefs: knowledge (with rationality is a inference). "coherence" notion. most general approach
Table 2: The Act Rationally paradigm compared to others
Compared to Advantage Think rationally More general, includes building more diverse, useful systems. Inference is a useful, but not necessary or sufficient mechanism for achieving rational behavior. Act like humans Human behavior includes quirky results of a poorly understood evolution.
3. Table 3: Contributions of other disciplines to AI
(in order by age of discipline)
Discipline Contributions Philosophy foundations of ontology and epistemology. theories of the mind, and behavior (means-end analysis). Mathematics formalization of algorithms, logic, probability. decision theory connects probability and utility theory. theories of incompleteness, and complexity (intractability, reduction, NP-completeness). Psychology AI grew most directly out of cognitive psychology. Three steps of agent processing: 1) translate stimulus into internal representation 2) manipulate to derive new internal representations 3) translate back into action. Computer hardware and software engineering Linguistics close relationship to AI in knowledge representation, and of course, NLP.
4. Lessons from the history of AI:
a. Early over-enthusiasm was due to a lack of understanding of time and space complexity. Scaling up is usually not linear: finding a solution in principle does not mean it can be found in practice.
b. Separating data (knowledge) from program (procedure) enhances clarity and flexibility of a system.
c. "Weak methods", which attempt to combine elementary reasoning steps to solve hard problems, usually don't work. Larger reasoning steps are required, too.
d. AI has not achieved human-like capabilities, but is building systems that work, that people pay money for.
Chapter 2: Intelligent Agents
1*. Agents perceive an environment, and act upon the environment towards achieving goals.
2. Selection of a rational action depends on (a coherence notion as opposed to omniscience):
a. An (external) performance measure that defines degree of success.
b. The agent's history of perceptions, the percept sequence.
c. The agent's knowledge of the environment.
d. The set of actions that are available.
3. Rationality under resource limitations (procedural rationality) is distinct from making the objectively rational choice (substantive rationality).
4. A spectrum of agents with increasingly sophisticated capabilities that build on each other:
1) simple reflex agents -- condition-action rules map between percepts and actions.
2) internal state agents -- condition-action rules map between state of the world and actions.
3) goal-based agents -- more flexible interpretation of the significance of environmental conditions.
4) utility-based agents -- differentiate between multiple goals, and/or ways to reach goals.
Also (may be combined with any of the agents above):
5) learning agents -- a learning element, separate from the performance element, modifies the performance element based on feedback.
5. Table 4: Properties of environments
Property (easy / hard) Description accessible / inaccessible Whether sensors perceive complete state of the environment. deterministic / non-deterministic Whether the next world state is a function of the current state and the agent's action. episodic / non-episodic Whether the current world state contains all relevant information (without previous world states). static / dynamic Whether the environment changes while the agent is deliberating. discrete / continuous Whether percepts and actions are distinct, clearly defined, separable.
Some questions to discuss
1. Does everybody agree that the paradigm of building intelligent agents is best for setting the AI research agenda?
2. In what ways is the intelligent agent paradigm embedded in a richer, deeper understanding?
My ideas on this:
functionalism, problem solving in complexity, modularity and stability
(see Herbert Simon, The Sciences of the Artificial, 1969).
3. An exercise:
a. identify where R & N discuss each type of agent (reflex, internal state, etc.).
b. identify where R & N discuss systems for operating in environments with each property.
c. consider the degree to which these categories are distinct, complete, and useful.