CSC242: Introduction to Artificial Intelligence Homework 1: AIMA Chapter 1 1. Describe in one sentence each the following approaches to creating artificial (computational) intelligences: (a) Acting humanly Solution: Focus on developing systems that can participate in activities (e.g., conversation) indistinguishably from humans, regardless of how such behavior is implemented internally. (b) Thinking humanly Solution: Develop intelligent systems by modeling how the human mind works, insofar as we understand this through psychological and neurological experiments. (c) Thinking rationally Solution: Use logical reasoning to compute “correct” conclusions (something humans do at least some of the time when behaving intelligently). (d) Acting rationally Solution: Develop systems that act rationally, in the sense of making appropriate choices (maximizing expected outcome) even when these cannot (always) be inferred deductively. Note that these are not mutually exclusive! 2. What makes the Turing Test a good test of intelligence? In what ways is it lacking? Solution: Pro: The TT requires language, knowledge, reasoning, learning, and so on, which are abilities that we associate with intelligence (intelligent behavior). The 1 TT places no constraints on the subjects of conversation, allowing the examiner to probe any aspect of human existence. Con: Original version (arguably) ignored perception and physical world. There is debate about whether being able to “fool” someone is sufficient as a test of intelligence. Duplicating human behavior may not necessarily be the way to go (classic example of this is pursing flight by copying birds vs. achieving powered flight using the aerodynamics of airfoils as in airplanes). 3. True or false: Cognitive Science combines computer models from AI with experimental techniques from Psychology, Neuroscience, and Linguistics to develop testable theories of the human mind. Solution: True 4. Describe the two main obstacles of the logical approach to AI. Are these obstacles only for logic? Solution: It is hard to put informal knowledge in formal (i.e., logical) terms. It is hard to do inference in practice (in fact, impossible in Gödel’s sense, even in principle). Since computer programs are formal languages, these are problems for any computational model of intelligence. 5. Briefly distinguish rational action from logical inference. Is perfect rationality feasible? Solution: An agent can’t always decide what to do by reasoning purely logically (deductively). That is, what to do is not always a deductive consequence of the agents’ beliefs. Nonetheless, the agent still has to choose what to do in order to achieve the best outcome. Perfect rationality is mathematically well-defined, but (a) it may not match human intuitions about intelligent behavior, and (b) it is computationally intractable anyway. We’ll see all these concepts in more detail later in the course. Page 2