Matches in SemOpenAlex for { <https://semopenalex.org/work/W38132835> ?p ?o ?g. }
Showing items 1 to 75 of
75
with 100 items per page.
- W38132835 abstract "This thesis defines a novel architecture for artificial agents that increases their autonomy while enhancing our trust. A value-driven agent consists of a reward function that is carefully aligned with human utility, and a set of skills encoded in a reactive language that embeds a learning algorithm. The agent learns how to maximize its reward from experience. The language (Icarus) extends current designs, and the learning algorithm (SHARSHA) generalizes existing algorithms. Taken together, they contribute a novel and powerful method for embedding domain knowledge in reinforcement learning problems. Since an Icarus agent can learn (via SHARSHA) to maximize its own reward and that reward is aligned with user concerns, the value-driven agent will resolve the best strategy within its ability to maximize user utility. We call this a ‘Be all you can be’ guarantee. It validates agent behavior in advance of learning, and increases our willingness to deploy highly autonomous systems. We conduct two experiments in a simulated vehicle control domain to demonstrate the benefit of the value-driven architecture. The first examines the effect of encoding domain knowledge in reinforcement learning problems. We show that additional distinctions about state improve performance but decrease learning rate, while additional plan structure can increase both learning rate and performance. The use of background knowledge also decreases plan size by three orders of magnitude relative to the expected formulation of our test problem. This suggests a qualitative change in the scope and efficacy of feasible learning applications. Our second experiment examines the benefit of the value-driven architecture for agent design. We show that different reward functions can generate qualitatively different behavior over the same set of skills. This supports the use of a novel design method: we can develop one fixed skill base for a given application area, and customize individual agents via programming by reward." @default.
- W38132835 created "2016-06-24" @default.
- W38132835 creator A5022164605 @default.
- W38132835 creator A5027008015 @default.
- W38132835 creator A5081275482 @default.
- W38132835 date "2001-01-01" @default.
- W38132835 modified "2023-09-27" @default.
- W38132835 title "Value-driven agents" @default.
- W38132835 hasPublicationYear "2001" @default.
- W38132835 type Work @default.
- W38132835 sameAs 38132835 @default.
- W38132835 citedByCount "10" @default.
- W38132835 crossrefType "book" @default.
- W38132835 hasAuthorship W38132835A5022164605 @default.
- W38132835 hasAuthorship W38132835A5027008015 @default.
- W38132835 hasAuthorship W38132835A5081275482 @default.
- W38132835 hasConcept C119857082 @default.
- W38132835 hasConcept C126255220 @default.
- W38132835 hasConcept C134306372 @default.
- W38132835 hasConcept C14036430 @default.
- W38132835 hasConcept C14646407 @default.
- W38132835 hasConcept C154945302 @default.
- W38132835 hasConcept C177264268 @default.
- W38132835 hasConcept C199360897 @default.
- W38132835 hasConcept C2776291640 @default.
- W38132835 hasConcept C33923547 @default.
- W38132835 hasConcept C36503486 @default.
- W38132835 hasConcept C41008148 @default.
- W38132835 hasConcept C41608201 @default.
- W38132835 hasConcept C78458016 @default.
- W38132835 hasConcept C86803240 @default.
- W38132835 hasConcept C97541855 @default.
- W38132835 hasConceptScore W38132835C119857082 @default.
- W38132835 hasConceptScore W38132835C126255220 @default.
- W38132835 hasConceptScore W38132835C134306372 @default.
- W38132835 hasConceptScore W38132835C14036430 @default.
- W38132835 hasConceptScore W38132835C14646407 @default.
- W38132835 hasConceptScore W38132835C154945302 @default.
- W38132835 hasConceptScore W38132835C177264268 @default.
- W38132835 hasConceptScore W38132835C199360897 @default.
- W38132835 hasConceptScore W38132835C2776291640 @default.
- W38132835 hasConceptScore W38132835C33923547 @default.
- W38132835 hasConceptScore W38132835C36503486 @default.
- W38132835 hasConceptScore W38132835C41008148 @default.
- W38132835 hasConceptScore W38132835C41608201 @default.
- W38132835 hasConceptScore W38132835C78458016 @default.
- W38132835 hasConceptScore W38132835C86803240 @default.
- W38132835 hasConceptScore W38132835C97541855 @default.
- W38132835 hasLocation W381328351 @default.
- W38132835 hasOpenAccess W38132835 @default.
- W38132835 hasPrimaryLocation W381328351 @default.
- W38132835 hasRelatedWork W2158150115 @default.
- W38132835 hasRelatedWork W2183087363 @default.
- W38132835 hasRelatedWork W2272929109 @default.
- W38132835 hasRelatedWork W2397581010 @default.
- W38132835 hasRelatedWork W2562252866 @default.
- W38132835 hasRelatedWork W2755614540 @default.
- W38132835 hasRelatedWork W2766302693 @default.
- W38132835 hasRelatedWork W2800367501 @default.
- W38132835 hasRelatedWork W2804924008 @default.
- W38132835 hasRelatedWork W2806124528 @default.
- W38132835 hasRelatedWork W2810531268 @default.
- W38132835 hasRelatedWork W2811111819 @default.
- W38132835 hasRelatedWork W2884970059 @default.
- W38132835 hasRelatedWork W2990485036 @default.
- W38132835 hasRelatedWork W2998219625 @default.
- W38132835 hasRelatedWork W3005053211 @default.
- W38132835 hasRelatedWork W3028389536 @default.
- W38132835 hasRelatedWork W3200996868 @default.
- W38132835 hasRelatedWork W76760840 @default.
- W38132835 hasRelatedWork W3095548673 @default.
- W38132835 isParatext "false" @default.
- W38132835 isRetracted "false" @default.
- W38132835 magId "38132835" @default.
- W38132835 workType "book" @default.