gary@rochester.UUCP (Gary Cottrell) (02/22/84)
From: Gary Cottrell <gary> University of Cottage Street Dept. of Dog Science 55 Cottage Street Rochester, New York 14608 SEMINAR Saturday, 25 February 1984 55 Cottage St. 9:00 p.m. Speaker Garrison W. Cottrell University of Cottage Street Topic "New Directions in Connectionist Dog Modeling" Further work has been done in the last year on the con- nectionist (spreading activation) formalism for modeling the generic dog (see Dog: A Canine Architecture, Cottrell 81; Toward Connectionist Dog Modeling, Cottrell 82). We extend the model to investigate sub-primate language use. This greatly simplifies the task of language study, since the language generation system consists of little more than S->"arf"|"rough"|"bark" thereby circumventing the transformational theories of Chom- sky and his gang of leftist thugs. Our first resuIt is that the language comprehended by the generic dog is S-> {"no"}{N|V} with no lexical items in common with the generation language. We can therefore study the comprehension system as totally independent of the generation system. In this impoverished domain, we may study lexical access to word information independent of any surrounding sentence struc- ture. In particular, we may study the well known effects of pragmatic context on comprehension. For example, the string "Call the elevator, JellyBean" when uttered in the context of "no elevator", causes the dog to jump on the nearest wall, whether it has a button or not.(1) Therefore, in this language domain, the appropriateness of the context is actu- ally computed by the speaker, while the hearer must make do with whatever resources are available to him at the time. The second result is that the dog has few internal resources available (little brain). We intend to put the model to a practical test by show- ing that previous "impoverished phoneme" naming verification experiments, which showed that the dog accepted "whiskey" as "biscuit" in the context of a biscuit, generalize to a forced choice paradigm, i.e., where whiskey is actually one of the choices, the generic dog will still choose the bis- cuit, showing that the problem is one of an inadequate world knowledge frame (no brain). The materials remaining from the test will be served to observers, in the interest of gaining introspective intuitions as to the nature of such a decimated world view. _____ 1. Such a result also shows that the lexicon of the dog may be collapsed from previous estimates of 20-30 words to a few modifiers of the current state. For example "get in the back" (of the car) when uttered in the office causes the animal to get up and move. Giving any credence to the look of confusion on his face is surely the result of anthropo- morphizing. ------------------------------------------------- This was a recent party announcement here. I was encouraged to share it with the world at large. Hope you enjoy it. gary cottrell
gary@rochester.UUCP (10/14/84)
From: Gary Cottrell <gary> University of Cottage Street Dept. of Dog Science 55 Cottage Street Rochester, New York 14608 SEMINAR Saturday, 20 October, 1984 55 Cottage St. 9:00 p.m. _M_o_d_e_l_l_i_n_g _t_h_e _I_n_t_e_n_t_i_o_n_a_l _B_e_h_a_v_i_o_r _o_f _t_h_e _D_o_g Garrison W. Cottrell Many of us, while out for a stroll, have had the experience of observing a dog trotting along, alone, obviously _g_o_i_n_g _s_o_m_e_w_h_e_r_e. This raises many questions, such as, "Where is he going?", "Why is he going there?", "Will it be more fun than where I'm going?", and so forth. Such questions motivate us to postulate the existence of (and hence the efficacy of further study of) the intentional behavior of the dog[1]. We propose a highly parallel, neurologically plausible model of dog behavior based on a connectionist (neural network) implementation of a subset of Reiter's (1980) Default Logic, as reported in Cottrell (1984). As outlined in that paper, there is a well specified mapping of default rules to connectionist network fragments that implement those rules, with the benefit that the network operates in real time by continuously updating the truth value of all predicates in parallel[2] (thus making Doyle's work, and perhaps Doyle himself, superfluous). Currently, the implementation only allows inference rules with one universally quantified variable. While inadequate for many purposes, we claim that this is all we need for dog modelling, since it appears that dogs can only think about one thing at a time anyway[3]. In this work we reinterpret ____________________ [1]Grembowitz (1982) proposed a model of the cat, but only handled the case of the cat tripping on catnip, cata- tonically staring at the wallpaper for hours with sporadic leaps into space. This simple behavior was elegantly modeled by the composition of only two standard UNIX calls, random(3c) and sleep(1). [2]The observant reader will recognize a certain similar- ity to British Motor Corporation's oft-lamented experiment of shoehorning an Austin Healey six cylinder engine into an MGB. Early results support the contention that our bastard child of a similar "marriage made in hell" will be more suc- cessful. [3]It is interesting to note that the set of things a dog can think about as noted in "Dog: A Canine Architecture", Cottrell (1981) may be _f_i_n_i_t_e and limited to food, squir- rels, and other dogs. Further, the dog we have studied ap- pears to only have three responses to other dogs, depending on their sex. Reiter's default inference rules as precondition-action behavior rules. An example behavioral rule is: Chase(x) ----------------------- Moving(x) : Squirrel(x) An English interpretation of this rule is, "if something is moving, and we don't have evidence that it isn't a squirrel, then chase it." This models the observed behavior of Jelly Bean chasing a paper bag. The real time behavior of our implementation captures his stopping when it turns out not to be a squirrel, since that blocks the inference of Chase(x), which then slowly decays, much as Jelly Bean slows to a confused halt. (As a simplifying assumption, we ignore his subsequent pretense of not having been chasing it at all.) Of course, we still have to determine whether there might still be some peanut butter and jelly in the bag, but this can be easily handled by the addition of more rules. Note that since we are building a model of behavior, the consequent of the rule is an action (Chase(x)), not an addition of the useless information Squirrel(x) to the already overtaxed knowledge base[4]. We have a grandiose long term research plan to model the entire mind of the dog, which will generate grant proposals _a_d _n_a_u_s_e_u_m. One of the new tools we plan to use in this research is the previously unnoticed ability to access the goal structures of the dog through measurement of tail wagging (for a discussion of some other aspects of tail wagging, including tail recursion, see "The Dog Papers", Benson & Sloan (1984)). We claim that tail wagging will do for dog modelling research what reaction times have done for psychology. For example, we can use this technique to assess the goal priorities of the dog. If we ask "do you want to go out?" we get a vigorous wagging response, whereas if we ask, "do you want to stay?", we get no tail wagging. Further, we can map out all of the levels of the system by studying the _t_i_m_e _c_o_u_r_s_e of the wagging behavior. Demonstrations of the the time course of the wagging response will be provided. ____________________ [4]As evidence that the knowledge base is already full, one only needs to note that when Squirrel(x) holds, and x climbs a tree, the dog repeatedly attempts to climb the tree by jumping on the trunk, even though this tactic has never been observed to succeed. (Copies of the troff source available on request) gary cottrell (allegra or seismo)!rochester!gary (USENET) gary@rochester (ARPA)