Susi Rule-Score Hierarchy Brainstorming

For Susi’s score system, we need a hierarchy to assign good score values to the rules. To do so we should develop a hierarchy to find an easy system that can be used to assign scores to new rules.

Please add your suggestions below, as you add your ideas we will change a score hierarchy suggestion below.

Preliminary Consideration: Patterns

We have two kinds of rules: such with patterns and others without. The meansing of such rules are:

with pattern(s):

  • (P+LR) variables in pattern should be used for retrieval in internal Susi’s log (reflection memory)
  • (P+IR) variables in pattern should be used for retrieval in internal databases
  • (P+ER) variables in pattern should be used for retrieval in external databases
  • (P+LS) variables in pattern should be stored in Susi’s memory to be used for reflection later
  • (P+IS) variables in pattern should be stored in internal databases to be used for retrieval later
  • (P+ES) variables in pattern should be stored in external databases to be used for retrieval later

without any pattern:

  • (P-D) default answers if no other rule applies
  • (P-O) overruling of rules which would apply, but should not

Secondary Consideration: Purpose

We have three kinds of purposes for Susi answers:

  • (/A) to answer on the users question
  • (/Q) to ask a question to the user in the context of an objective within Susi’s planning to do a conversation
  • (/R) to answer on an answer of the user within the planning of Susi to do a conversation. It appears clear that answers in the context of a Susi conversation strategy should have higher priority.

Combinations of Pattern and Purpose Considerations:

To combine the various Pattern and Purpose types, we write the abbreviations of these items together. For example, we want to answer on a question of the user “Are you happy” with “Yes!, Are you happy as well?” which would be an rule of type P-O/Q. The combination of the both consideration types give 8×3=24 possibilities.

Score Hierarchy

I believe there should be
– score(R) > score(Q) > score(A):
to do a steering of conversations within a conversation plan.
– score(P-O) > score(P+?) > score(P-D):
overruling of pattern-directed answers and default answers in case of pattern fail
– score(P+?S) > score(P+?R):
storing of information (= learning) is more important than answering
– score(P+L?) > score(P+I?) > score(P+E?):
using local information is primary above external information. Reflection is most important.

This produces the following order (with decreasing score, first line has highest sore):

– Overruling of patterns:
– R/P-O
– Q/P-O
– A/P-O

– Answer on an Answer of the user using patterns, possibly learning, otherwise retrieving data
– R/P+LS
– R/P+IS
– R/P+ES
– R/P+LR
– R/P+IR
– R/P+ER

– Asking the user a question with the purpose of learning with the users answer
– Q/P+LS
– Q/P+IS
– Q/P+ES
– Q/P+LR
– Q/P+IR
– Q/P+ER

– Just giving an answer to the question of the user
– A/P+LS
– A/P+IS
– A/P+ES
– A/P+LR
– A/P+IR
– A/P+ER

– Fail-over if no other rule apply to just answer anything, but try to start a new conversation
– R/P-D
– Q/P-D
– A/P-D

Susi Rule-Score Hierarchy Brainstorming