COMEX: A Commodities Support System

COMEX: A COMMODITIES SUPPORT SYSTEM.

JIM STANSFIELD M.I.T. A.l. LAB

We are developing an intelligent support system, COMEX, to manage the data that a commodities analyst receives in agricultural reports. COMEX must have the following capabilities: 1) finding relevant facts, 2) determining implications, 3) checking reliabililies, 4) monitoring, 5) alerting, 6) reporting, 7) answering queries. The research involves natural language, representation, common sense reasoning, qualitative simulation an~ user modelling. Here, I address representation and reasoning by. describing COMEX-0 a working prototype of a module of COMEX.

CQMEX-0 uses frames (Minsky, 1975), (Goldstein & Roberts;. 1977) for representing knowledge. About twenty types of event are represented, together with crop-states, weather, commodities, quantities and prices. Frame structures are interwoven with procedures. COMEX-0 builds frame slrl!clures by means of a formal dialogue with a user : Each frame has procedures for requesting information to instantiate its slol.s. Procedures attached to slots check that responses satisfy the requirements of the slot. Unsatisfied requirements complain giving a reason for unacceplability.

Represenlinc events.

EVENTS relate objects and actors around several central actions. Represented events include TRANSFER, GIVE, BUY; SELL, TRANSPORT, CHANGE, TAKE, INCREASE, DECREASE, CAUSE, EXCHANGE, MOVE, EXPORT and IMPORT. COMEX-0 forms new event representations in two ways, INSTANTIATION of an event and AGGREGATION of several events.

Instantiation is illustrated by MOVING which is a kind of CHANGING. A pseudo-engl ish version of part of the MOVING frame ·is shown in figure 1.

MOVING has an extra SUBJECT slot for the thing moved and ha.s requirements that give ·more detai.led specifications of the slot values. It inherits time information from CHANGING which, in turn, inherits from EVENT. If-added and requirement procedures, as well as assertions, are inherited.

Aggregation is exemplified by EXCHANGING which has two parts, both of them TRANSFERINGS, which occur at the same lime. When we instantiate EXCHANGING, mapping commands use definiiions to construct the subcomponents so · as to represent that ACTOR! lransfered OBJECT! to ACTOR2 and ACTOR2 transfered OBJECT2 to ACTOR!. This mapping makes inferences, since assertions about TRANSFER lNG have become · available through expanding the definition of EXCHANGING. Mappings lake account of constraints among slots of the sub-components. They are like Merlin mappings (Moore and Newell, 1973) but are used differently. Parts oi a defined structure which use different frames but contain common elements are mapped into each .other.

A more complex event that COMEX-0 can build is· EXPORTING. This includes a TRADING and a TRANSPORTING as subevents .. The TRADING is an instance of EXCHANGING and has two TRANSFERs. The TRANSPORTING is a CAUSING to MOVE. The MOVE is a CHANGE.

Generic frames have instructions for building their Instances. These specify the parts, their arrangement and the constraints between them. Controlling the effects of constraints is the hardest part . The power of COMEX-O's representation scheme depends on further solutions to this.

lnferencing.

lnferencing is done in three ways: 1) definition expansion 2) constraints 3) matching. Expansion has been dealt with already.

Constraints are exemplified by SUPPL V-DEMAND frames which specify amounts of wheat in various sectors of the market. Constraints between slots ·of this frame (e.g. old- supply ,. production · new-supply) are represented as if-added theorems attached to the slots. No search is needed to invoke these.

Matching is a third inference .mechanism. When two matching structures represent one object, each may have .some unique information. When merged, the information will all be directly related. COMEX-0 can match arbitrary frame structures. The requirements restricting values of a slot are used when the slot is matched asainsl a structure. Local matching functions specify how to match against specific frame types. The lest for matchability of two frames is whether a generic description of each fils the other.

Conclusion

COMEX-0 is being extended to include a qualitative simulation of wheat growth. By this we mean it deals with descriptive and incomplete information. as well as exact quantitative data. The event representation mechanism is being use.d to set up qualitative process scenarios such as rain and snow falling on, being absorbed by and running off soil.

References. ·

Goldstein, J.P. & Roberts, R.B. 1977, NUDGE: a knowledge based scheduling program. MIT AI Lab memo 405.

Minsky, M. 1975, A framework for representing knowledge. in The psychology of computer vision. Ed. Winston, McGraw-Hill.

Moore, J. and Newell, A. 1973, aiHow can MERLIN understand? in Knowledge and Cognition. Ed. L. Gregg, Erlbaum.