Kaelbling, Leslie P. | Engineering Papers

NASA NTRS1993

Acting to gain information

This report is concerned with agents that act to gain information. In previous work, we developed agent models combining qualitative modeling with real-time control. That work, however, focused primarily on actions that affect physical states of the environment. The current study extends that work by explicitly considering problems of active information-gathering and by exploring specialized aspects of information-gathering in computational perception, learning, and language. In our theoretical investigations, we analyzed agents into their perceptual and action components and identified these with elements of a state-machine model of control. The mathematical properties of each was developed in isolation and interactions were then studied. We considered the complexity dimension and the uncertainty dimension and related these to intelligent-agent design issues. We also explored active information gathering in visual processing. Working within the active vision paradigm, we developed a concept of 'minimal meaningful measurements' suitable for demand-driven vision. We then developed and tested an architecture for ongoing recognition and interpretation of visual information. In the area of information gathering through learning, we explored techniques for coping with combinatorial complexity. We also explored information gathering through explicit linguistic action by considering the nature of conversational rules, coordination, and situated communication behavior.

Rosenchein, Stanley J.↗

NASA NTRS1991

Collected notes from the Benchmarks and Metrics Workshop

In recent years there has been a proliferation of proposals in the artificial intelligence (AI) literature for integrated agent architectures. Each architecture offers an approach to the general problem of constructing an integrated agent. Unfortunately, the ways in which one architecture might be considered better than another are not always clear. There has been a growing realization that many of the positive and negative aspects of an architecture become apparent only when experimental evaluation is performed and that to progress as a discipline, we must develop rigorous experimental methods. In addition to the intrinsic intellectual interest of experimentation, rigorous performance evaluation of systems is also a crucial practical concern to our research sponsors. DARPA, NASA, and AFOSR (among others) are actively searching for better ways of experimentally evaluating alternative approaches to building intelligent agents. One tool for experimental evaluation involves testing systems on benchmark tasks in order to assess their relative performance. As part of a joint DARPA and NASA funded project, NASA-Ames and Teleos Research are carrying out a research effort to establish a set of benchmark tasks and evaluation metrics by which the performance of agent architectures may be determined. As part of this project, we held a workshop on Benchmarks and Metrics at the NASA Ames Research Center on June 25, 1990. The objective of the workshop was to foster early discussion on this important topic. We did not achieve a consensus, nor did we expect to. Collected here is some of the information that was exchanged at the workshop. Given here is an outline of the workshop, a list of the participants, notes taken on the white-board during open discussions, position papers/notes from some participants, and copies of slides used in the presentations.

Drummond, Mark E.↗