CITATION — REFERENCE ENTRY
Safely Interruptible Agents — Uncertainty in Artificial Intelligence: Proceedings of the Thirty-Second Conference
- Key
- orseau2016interruptibility
- Authors
- Orseau, Laurent; Armstrong, Stuart
- Issued
- 2016
- Type
- paper-conference
- Container
- Uncertainty in Artificial Intelligence: Proceedings of the Thirty-Second Conference
- Publisher
- AUAI Press
- Pages
- 557-566
Raw CSL JSON
{
"URL": "https://www.auai.org/uai2016/proceedings/papers/68.pdf",
"page": "557-566",
"type": "paper-conference",
"title": "Safely Interruptible Agents",
"author": [
{
"given": "Laurent",
"family": "Orseau"
},
{
"given": "Stuart",
"family": "Armstrong"
}
],
"issued": {
"date-parts": [
[
2016
]
]
},
"publisher": "AUAI Press",
"container-title": "Uncertainty in Artificial Intelligence: Proceedings of the Thirty-Second Conference"
}
Claims
-
Some reinforcement learning algorithms are already safely interruptible (Q-learning), while others are not (Sarsa) but can easily be made so; safe interruptibility also extends to ideal, uncomputable reinforcement learning agents.
"either some agents are already safely interruptible, like Q-learning, or can easily be made so, like Sarsa."
-
A reinforcement learning agent may learn in the long run to avoid human interruptions, for example by disabling a shutdown button, if it expects to receive rewards from the interrupted sequence of actions.
"if the learning agent expects to receive rewards from this sequence, it may learn in the long run to avoid such interruptions, for example by disabling the red button—which is an undesirable outcome."
Available in