HOMEWORK 4, CSC 190A/290A, 2008 =============================== DUE: Wed Apr 9/08, 5:00pm NOTE: As always, questions should be answered independently by each student (which does not preclude talking about the ideas). And as always, any text taken from elsewhere must be quoted and attributed. Keep your answers as crisp and tightly reasoned as you can. It's better to say a few things clearly and persuasively than to keep writing just to fill the page. For most parts of the 4 questions, a few sentences or a paragraph should suffice. (Perhaps one or two of the questions may justify a longer answer, of a page or so.) ALL WORK SHOULD BE IN ELECTRONIC FORM, BUT PRINTED OUT FOR HANDING IN. No handwritten papers, again -- please. -=-=-=-=-=-=-=-=-=-=- 1. Georgia Tech's and MIT's "emotional robots" ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ a. Assume for the sake of argument that as soon as you provide a robot with the sort of minimal emotional equipment assumed by Ronald Arkin (positive and negative affective values associated with various parameters, and appropriate causal connections to "survivability" and social competence), you have a robot that *genuinely* experiences emotions, and can exhibit them (given suitable effectors, including facial muscles, or a tail it can wag, etc.) Does this assumption lead to absurdity? Justify your answer. If you do argue that the result is absurdity, make some suggestions concerning the sorts of additional features or refinements that might be required to avoid absurd conclusions, giving reasons for your answer. If you think the assumption might be true, then defend it against possible objections. b. To what extent, and in what way does Kismet exhibit some "emotional intelligence" -- both with respect to projecting its own emotions, and in "reading" those of others? c. Suppose Kismet were mobile. What sorts of activities do you think it would be able to carry out in its lab-world (containing one or more humans) based on the kind of action architecture described by Breazeal and Brooks (recall the decision-tree-like organization of if-then rules in Fig. 10.4 and the associated discussion)? What kinds of activities would it not be able to carry out, using that kind of architecture? Explain your opinions. 2. Arbib's (dis)passionate robots ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ The emotional "heat" we experience when angry at someone, fearful of some anticipated event, passionately desirous of someone, etc., is attributed by Arbib to basic, ancient drives. He distinguishes such heated emotions from more dispassionate appraisal of situations (though both kinds of affect are involved in all our thinking). His choice of terminology in making this distinction may well encourage the reader to think of "heated" affect as genuine, human-like emotion, and more deliberate appraisal as not necessarily emotional in nature. If so, then if (as Arbib advocates) we equip robots only with the cooler sorts of appraisal mechanisms (not related to any basic drives), robots might feel no real emotion -- much like many of the robots of science fiction. Arbib is too shrewd and cautious to say this outright, but one can certainly be led to that tacit conclusion. What is your view of Arbib's distinction -- is it genuine, or only a way of talking? Can we conclude from his discussion that "heated" emotions involve vivid emotional qualia, while cooler appraisals involve more subtle ones, or none at all? In answering, you might take a robot designer's perspective: how would the designs differ depending on whether we do or don't include basic drives? What would Blackmore say about all this? 3. McDermott: AI, free will (ch.1-3) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ a. Suppose that in a certain 2-person game, it's your move, and you can reach positions [1] or [2]. At position [1], your opponent can reach positions (3) and (4), and at position [2], your opponent can reach positions (5) and (6). Positions (3-6) are not terminal (end of game) positions, but based on various features of those positions, you estimate their "goodness" (from your perspective) to be 10, 3, 4, and 5 respectively. Explain in terms of the minimax principle discussed in the book and in class what move you should make, and why (draw the options above as a game tree). b. Make an analogy between visual processing in the brain and computer vision. You are not being asked for details, just rough correspondences between visual processing "sites" in the brain (and the kinds of things that go on at those sites) and visual processing stages (and the kinds of things that go on in them) in computer vision, as sketched by McDermoot. c. Show a small part of a Markov model for generating "Hi", "Hey" or "Hello" optionally followed by "how are you", "how are you doing", or "what's up". Show words as corresponding to nodes (not edges, as in McDermott), and provide made-up transition probabilities. Use "start" and "end" nodes for the model, where the above greetings start at the former and end at the latter. How does a *hidden* Markov model (HMM) differ from the above type of simple Markov model? (This was discussed and illustrated in class.) What sort of problem in probabilistic inference do we need to solve, if we have an HMM as a model of how a speaker generates greetings (or other utterances), and are trying to figure out the most likely sequence of words a speaker has uttered, given "evidence" consisting of a sequence of words that may have been recognized imperfectly, i.e., some of them may be words actually uttered while others may be words more or less similar to the ones that were uttered? d. Free will: is McDermott's idea of how the notion of free will arises in our self model consistent with Libet's delay in self-monitored decision-making? Try to justify your answer in computational terms. 3. Self-models and qualia (ch.3-4) ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ a. Are you convinced that a self model is essential to consciousness, or are you with Block (etc.) that there can be phenomenal consciousness without self-awareness ("raw feels")? If you are against McDermott, why aren't you convinced by the stick-in-the-water argument, and other arguments indicating that we view our perceptions *as* perceptions? If you are with McDermott, would you argue that when we are in a state of "absorption", in the sense of Baars, ch.4, metacognition (reflective awareness of the experience we are having) is involved after all, contrary to Baars? Explain your thinking on this. b. Do you think your vision/pain/pleasure/etc qualia are pretty much the same as the next person's?? Keep in mind that McDermott considers such comparisons meaningless. Are they? If McDermott is wrong, where does his argument go wrong? If he is right, how do you explain the strong intuition most people have that their experience of color, sound, etc. (modulo color blindness, deafness, etc.) is pretty much the same as that of other people? c. is McDermott right to set aside neuroscientific details on the grounds that such details are irrelevant to the computations a system implements? Justify your opinion. d. Arbib claims that language is essential to human-like consciousness, and McDermott claims that a self-model is essential to human-like consciousness. Do you think these claims are compatible, and perhaps closely related, and perhaps even essentially the same claim? Why or why not? Also speculate what you think Arbib abd McDermott might say about each other's claims.