next up previous
Next: Learning similarity from Up: Learning to Adapt Previous: Learning to Adapt

Effects of learning response plan cases and adaptation cases

  We predicted that the acquisition of more specific knowledge about useful adaptations would decrease adaptation effort. To obtain initial indications of the effects, we performed ablation tests of the system with our test examples. The memory included nodes for 870 concepts; the case library included 6 disaster response plan cases, and the test examples involved performing a total of 30 adaptations to develop response plans for 5 stories. Stored cases and new stories were based on the Clarinet News Service newswire and the INvironment newsletter for air quality consultants.

The experiments examined the DIAL system's performance on a static set of examples across a variety of learning settings. Because most of the system's adaptation cost comes from memory search, efficiency was estimated by the number of memory operations performed and the number of memory nodes visited. In each case, lower values suggest less effort expended. Table 1 shows the average, maximum and minimum number of operations which are applied during each type of learning.

 

Memory Ops Nodes visited
Avg Max Min Avg Max Min
Using ``local search'' to find needed information
1. No learning 103 226 7 53 114 4
2. Plan learning only 80 214 7 41 108 4
3. Adaptation learning only 68 181 4 40 92 3
4. Plan + Adaptation Learning 66 181 4 39 92 3
Using multiple strategies to find needed information
5. Plan Learning 548 812 42 56 83 5
6. Plan + Adaptation Learning 59 312 1 26 50 1

Table 1:   Average, maximum and minimum effort expended adapting the five sample cases.

The results suggest that adaptation learning produced slightly better results than normal case learning, although the combination of the two produced insignificant additional benefit. When given a larger set of search primitives from which to learn, adaptation learning reduced the cost of search in all tests. We plan to follow up on these initial data by performing a more controlled analysis of the effects of learning for a larger set of problem examples, and also to examine the potential utility problem [Minton1988] as the number of adaptation cases grows.



Andrew Kinley
Thu Apr 4 10:19:36 EST 1996