Using Qualitative Preferences to Guide Schedule Optimization
As robots and computers become more capable, we see more mixed human-robot- computer teams. One strategy for coordinating mixed teams is standardizing priorities across agents. However, the encoding of priorities may lead to discrepancies in interpretation across different types of agents. For examp...
| Main Author: | |
|---|---|
| Other Authors: | |
| Format: | Thesis |
| Published: |
Massachusetts Institute of Technology
2023
|
| Online Access: | https://hdl.handle.net/1721.1/150116 |
| _version_ | 1811072580339105792 |
|---|---|
| author | Wells, Tesla |
| author2 | Williams, Brian C. |
| author_facet | Williams, Brian C. Wells, Tesla |
| author_sort | Wells, Tesla |
| collection | MIT |
| description | As robots and computers become more capable, we see more mixed human-robot- computer teams. One strategy for coordinating mixed teams is standardizing priorities across agents. However, the encoding of priorities may lead to discrepancies in interpretation across different types of agents. For example, a machine may have difficulty generating a plan that incorporates the subtleties of a natural language description. A human, given a purely quantitative encoding, may have difficulty intuiting trends or generating a plan without a computation aid. These discrepancies in priority interpretation may lead to a plan that only one class of agent deems optimal. In this case of distributed planning, differing interpretations may lead to significant differences in the plans produced by different classes of agents. What’s more, low fidelity interpretation of priorities can prohibit model correction or refinement. A robot may not recognize its learnt-physics-model conflicts with underlying assumptions in a natural language description of priorities. For humans, it is difficult to identify incorrect or unrefined priority models from the same streams of numerical relations computers find useful.
Most strategies currently in use for bridging this gap involve machines learning human preferences from large data sets or require labor-intensive custom utility encodings to be written, explained, and revised by a trained expert. The former often homogenizes models of human preferences and fails to incorporate corrections accurately, efficiently, and in context. The latter prohibits the usage of robots or computer aids in casual or dynamic settings without the presence/supervision of a human trained in working with the system. In both settings, this inhibits average humans from having personalized, human-computer or human-robot interactions.
In this thesis, we attempt to improve human-robot interactions by encoding utility as a series of qualitative, Ceteris Paribus preference statements over a design space. We posit this formalism is both readily understood by human agents and easily reasoned over by machines. Previous work establishes the ability to compute machine readable utility functions from said statements by efficiently generating topological orderings. We detail our implementation of a machine “agent” who uses said utility function to compute optimal schedules to Conditional Temporal Problem problems for a mixed-agent team. We then build off of the procedure for generating utility functions to generate admissible heuristics that increase performance.
We show this encoding enables us to explain which human preferences differentiate feasibly scheduled, high-utility plans. We present a suit of algorithms capable of explaining model behavior according to five different relevance standards. Additionally, we construct and algorithm for identifying where additional preference specification would resolve assumptions used in underspecified areas of the model. These explanations not only improve human understanding, but also facilitate identification of inaccuracies in the machine’s utility model. We build off our explanations to presents options for model-repair. We show preference-addition and preference-relaxation informed by explanation results in specific, targeted plan changes. |
| first_indexed | 2024-09-23T09:08:17Z |
| format | Thesis |
| id | mit-1721.1/150116 |
| institution | Massachusetts Institute of Technology |
| last_indexed | 2024-09-23T09:08:17Z |
| publishDate | 2023 |
| publisher | Massachusetts Institute of Technology |
| record_format | dspace |
| spelling | mit-1721.1/1501162023-04-01T03:56:54Z Using Qualitative Preferences to Guide Schedule Optimization Wells, Tesla Williams, Brian C. Massachusetts Institute of Technology. Department of Aeronautics and Astronautics As robots and computers become more capable, we see more mixed human-robot- computer teams. One strategy for coordinating mixed teams is standardizing priorities across agents. However, the encoding of priorities may lead to discrepancies in interpretation across different types of agents. For example, a machine may have difficulty generating a plan that incorporates the subtleties of a natural language description. A human, given a purely quantitative encoding, may have difficulty intuiting trends or generating a plan without a computation aid. These discrepancies in priority interpretation may lead to a plan that only one class of agent deems optimal. In this case of distributed planning, differing interpretations may lead to significant differences in the plans produced by different classes of agents. What’s more, low fidelity interpretation of priorities can prohibit model correction or refinement. A robot may not recognize its learnt-physics-model conflicts with underlying assumptions in a natural language description of priorities. For humans, it is difficult to identify incorrect or unrefined priority models from the same streams of numerical relations computers find useful. Most strategies currently in use for bridging this gap involve machines learning human preferences from large data sets or require labor-intensive custom utility encodings to be written, explained, and revised by a trained expert. The former often homogenizes models of human preferences and fails to incorporate corrections accurately, efficiently, and in context. The latter prohibits the usage of robots or computer aids in casual or dynamic settings without the presence/supervision of a human trained in working with the system. In both settings, this inhibits average humans from having personalized, human-computer or human-robot interactions. In this thesis, we attempt to improve human-robot interactions by encoding utility as a series of qualitative, Ceteris Paribus preference statements over a design space. We posit this formalism is both readily understood by human agents and easily reasoned over by machines. Previous work establishes the ability to compute machine readable utility functions from said statements by efficiently generating topological orderings. We detail our implementation of a machine “agent” who uses said utility function to compute optimal schedules to Conditional Temporal Problem problems for a mixed-agent team. We then build off of the procedure for generating utility functions to generate admissible heuristics that increase performance. We show this encoding enables us to explain which human preferences differentiate feasibly scheduled, high-utility plans. We present a suit of algorithms capable of explaining model behavior according to five different relevance standards. Additionally, we construct and algorithm for identifying where additional preference specification would resolve assumptions used in underspecified areas of the model. These explanations not only improve human understanding, but also facilitate identification of inaccuracies in the machine’s utility model. We build off our explanations to presents options for model-repair. We show preference-addition and preference-relaxation informed by explanation results in specific, targeted plan changes. S.M. 2023-03-31T14:33:25Z 2023-03-31T14:33:25Z 2023-02 2023-02-15T14:06:26.283Z Thesis https://hdl.handle.net/1721.1/150116 0000-0003-2310-0795 In Copyright - Educational Use Permitted Copyright MIT http://rightsstatements.org/page/InC-EDU/1.0/ application/pdf Massachusetts Institute of Technology |
| spellingShingle | Wells, Tesla Using Qualitative Preferences to Guide Schedule Optimization |
| title | Using Qualitative Preferences to Guide Schedule Optimization |
| title_full | Using Qualitative Preferences to Guide Schedule Optimization |
| title_fullStr | Using Qualitative Preferences to Guide Schedule Optimization |
| title_full_unstemmed | Using Qualitative Preferences to Guide Schedule Optimization |
| title_short | Using Qualitative Preferences to Guide Schedule Optimization |
| title_sort | using qualitative preferences to guide schedule optimization |
| url | https://hdl.handle.net/1721.1/150116 |
| work_keys_str_mv | AT wellstesla usingqualitativepreferencestoguidescheduleoptimization |