Problems in Building an Instructable Production System

The Inst ructable Production System project is e x p l o r i n g the incremental g rowth propert ies of production sys tems (PSs) by construct ing a generally intelligent p rob lem-so l v i ng system by gradual (external) instruction. The de f in i t i on of PS and our current architecture are given e l s e w h e r e in this volume (Newell, 1977; Forgy and McDermot t , 1977). The present task domain is an abstract job shop, in which f inished goods are made from raw mater ia ls . We start w i th a Kernel (a small PS of about 200 p roduc t i ons ) which has the basic capabilities to grow by i ns t r uc t i on : (1) process a restr ic ted natural language; (2) f o rm produc t ions f rom its input; (3) impose PS control conven t ions on them; and (4) per form basic manipulations in its env i ronment (Rychener & Newell, 1977). We take the basic computat ional and representat ional adequacy of PSs for AI p rograms as establ ished. This short note presents some immediate diff icult ies we expect to encounter. These derive from the ins t ruc t iona l s i tuat ion: (1) The instructor can observe the sys tem in the environment and can communicate with it f r e e l y , but cannot examine its internal structure direct ly. (?) In te rac t ion w i t h the system is in an external language, analogous to natural language. (3) The initiative for i n te rac t i on is mixed. (4) Instruct ion may be on any topic: speci f ic tasks, general proper t ies of tasks, the language of communicat ion, possible e r ro rs , how to plan and explore, etc. (5) Knowledge and system structure gained through i ns t ruc t i on accumulates over the life of the system. Our cu r ren t approach uses m^ns-e^Kls analysis as the basic, phi losophy of both problem-solving and ins t ruc t i on . Goals are symbol structures in Working Memory that descr ibe desired states and processing e n t i r e l y t h rough the means and tests. Means are encoded as p roduc t ions that recognize goals and assert subgoals whose sat is fact ion wil l achieve the goals. Tests are encoded as product ions that recognize the conditions of sa t i s fac t ion of the goal. The means productions form a means-ends ne twork of goals. Instruction consists of e l abo ra t i ng the nodes of this network as required by a task. Now for the dif f icult ies on the immediate horizon: 1. Contact : How can initial contact be made wi th ex is t ing knowledge that might be relevant to the task at hand, wh ich is not part of the means-ends network de l i be ra te l y c rea ted by instruct ion for the task? Use of the data acquired through experience is required in any in te l l igent agent. Once detected, much processing can be spent on d iscover ing relevance, but initial contact may be e x t r e m e l y d i f f icu l t . Any general intelligent system wil l have too much knowledge to consider exhaustively. PS a rch i tec tu res exploi t this by stor ing all knowledge as p roduc t i ons which are evoked only if their conditions "see themse lves" in the Working Memory. But means p roduc t ions are acquired in specific contexts and their cond i t ions become keyed to specific goals and task fea tu res . One approach might be to generate variations of c u r r e n t Work ing Memory goals and data until something is evoked . 2. Incoherency: The PS may be essentially incoherent in descr ib ing its situation and diff icult ies to an ex te rna l inst ructor . The means-ends network helps (by p rov id i ng the same level of explanatory capabil i ty as in cu r ren t exper t systems), but is not sufficient. E.g., descr ibe what went wrong f rom the debris left in Working Memory . The PS's diagnostic and explanatory capabilit ies are expandable by instruct ion, but it is current ly unclear how this wi l l work. 3. Means-ends analysis eff iciency: Initial instruct ion p roduces a more elaborate network than is necessary. The ins t ruc to r uses numerous intermediate goals, both to make his ins t ruc t ion sequences easier to generate and to allow complex procedures to be taught at all; miscommunication leads to a pa tchwork of variant procedures; the PS uses a goa l -encumbered monitor ing mode of operat ion insert ing s u p e r v i s o r y goals and processes; etc. There are three modes of va ry ing eff ic iencies: a compiled, efficient mode; o r d i n a r y ins t ruc t ion mode; and the monitoring mode. Our p r ima ry concern is t ransforming from the ordinary form to the compi led fo rm, while maintaining a capabil ity to rever t back to the other two in debugging situations. This may not be at tainable simply through instruct ion; architectural modi f ica t ion may be required. 4. Uti l iz ing distant ly related knowledge: Knowledge about o ther tasks is imperfect for a given task; it is also embedded in methods and encoded in representat ions c rea ted for (and local to) the distant task. All these aspects cause d i f f icu l ty , even if contact is made (per 1). A clear symptom wi l l be repet i t ive instruction to cover minor task var ia t ions. One approach is to avoid the dif f iculty by adopt ing un i form conventions for encoding. We think this won ' t work . We favor attempting to map methods to methods (and representat ions to representations), using ideas f rom Mer l in (Moore & Newell, 1973).