Johannes Treutlein & Caspar Oesterheld (both researchers at the Foundational Research Institute), EA Global X Berlin 2017, Oct 14/15 2017

1. Johannes Treutlein
Foundational Research Institute
Decision theory research
at FRI

2. Johannes Treutlein
Foundational Research Institute
A wager for evidential decision
theory

3. Altruistic Newcomb problem
3
Ω
?
one
wish
predicts one-boxing:
two wishes
predicts two-boxing:
nothing

4. Altruistic Newcomb problem
4
S1 S2
A1 2 0
A2 3 1
● A1: One-box; A2: Two-box
● S1: opaque box contains two wishes; S2: opaque box empty

5. Evidential decision theory
5

6. Causal decision theory
6

7. Meta decision theory
7
(Nozick 1993; MacAskill 2016)

8. 8
Altruistic Newcomb problem in a large
universe
Ω
Ω
Ω
Ω
Ω
Ω
Ω

9. Altruistic Newcomb problem in a large
universe
9

10. EDT Wager
10
● Large universe
● Caring about the gains of our copies
● Non-zero credence in EDT
● Meta decision theory
Wager for evidential decision theory (and all other theories that
take impact of copies into account)

11. Relevance
11
● AI Safety
● Macrostrategy
● Multiverse-wide superrationality (Oesterheld 2017a)

12. Caspar Oesterheld
Foundational Research Institute
Decision theory and approval-
directed agents

13. Implementing decision theories in AIs
13
• Two problems of decision theory in AI safety:
• What is the right decision theory for an AI?
• How do we implement decision theories in AI?
• Decision theory not explicit in AI architecture
• Example: Doing what has worked well in the past (Oesterheld
2017b)
• Exception: Gödel machine (Schmidhuber 2006)

14. Approval-directed agency
14
(Christiano 2014)

15. Two decision theories
15

16. Two decision theories
16

17. Example
17

18. Two decision theories
18

19. Example
19

20. 20
In the paper…
If overseer only looks at the world, the agent’s DT is
decisive.
If overseer only looks at the agent’s action, the
overseer’s DT is decisive.

21. Presentation title
John Smith | Head of Department 28.06.2016
Subtitle or caption
Thank you.
{johannes,caspar}@foundational-research.org

22. References
22
• Ahmed, A. (2014): Evidence, Decision and Causality. Cambridge University Press.
• Almond, P. (2010): On Causation and Correlation. Part 2: Implications of Evidential
Decision Theory. https://casparoesterheld.files.wordpress.com/2017/03/
correlation2.pdf
• Bostrom, N. (2014b): Superintelligence: Paths, Dangers, Strategies. Oxford
University Press.
• Christiano, P. (2014): Model-free decisions. https://ai-alignment.com/model-free-
decisions-6e6609f5d99e
• MacAskill, W. (2016): Smokers, Psychos, and Decision-Theoretic Uncertainty. The
Journal of Philosophy
• Nozick, R. (1993): The Nature of Rationality. Princeton: Princeton University Press

23. References
23
• Oesterheld, C. (2017b): Doing what has worked well in the past leads to evidential
decision theory. https://casparoesterheld.files.wordpress.com/2017/09/learningdt.pdf
• Oesterheld, C. (2017a): Multiverse-wide Cooperation via Correlated Decision
Making. https://foundational-research.org/files/Multiverse-wide-Cooperation-via-
Correlated-Decision-Making.pdf
• Schmidhuber, J. (2006): Gödel Machines: Self-Referential Universal Problem Solvers
Making Provably Optimal Self-Improvements. ftp://ftp.idsia.ch/pub/juergen/gm6.pdf
• Soares, N. and Fallenstein, B. (2014a): Aligning Superintelligence with Human
Interests: A Technical Research Agenda. MIRI Tech. rep. 2014-8. https://
intelligence.org/files/TechnicalAgenda.pdf
• Soares, N. and Fallenstein, B. (2014b): Toward Idealized Decision Theory. MIRI
Tech. rep. 2014-7. https://arxiv.org/abs/1507.01986
• Soares and Levinstein (2017): Cheating Death in Damascus. https://intelligence.org/
files/DeathInDamascus.pdf