A Dutch book against sleeping beauties who are evidential decision theorists

Vincent ConitzerSynth 2015 · OpenReview

Abstract: In the context of the Sleeping Beauty problem, it has been argued that so-called “halfers” can avoid Dutch book arguments by adopting evidential decision theory (EDT). This paper introduces a new variant of the Sleeping Beauty problem—the White-Black-Grey (WBG) variant—to demonstrate that EDT agents fall prey to a Dutch book regardless of whether they are halfers or thirders. The argument hinges on the idea that an action provides evidence for what an agent would do at other decision points with symmetric, even if distinguishable, information.

The Sleeping Beauty Problem and Decision Theory

The Sleeping Beauty problem is a classic paradox of self-locating belief. Beauty is put to sleep on Sunday. A fair coin is tossed: if Heads, she is awoken on Monday; if Tails, she is awoken on Monday and Tuesday (with her memory wiped between). When she wakes, she does not know the day. "Halfers" argue her credence in Heads should be 1/2, while "Thirders" argue it should be 1/3.

Philosophers use "Dutch books"—sets of bets that guarantee a loss—to test the rationality of these positions. While previous research suggested that agents using Evidential Decision Theory (EDT) could avoid such traps, this paper introduces a variant that proves otherwise.

The White-Black-Grey (WBG) Variant

To challenge the immunity of EDT, Conitzer proposes the WBG variant. In this setup, two coins are tossed. Coin 1 (White/Black) determines the color of the room on Monday. Coin 2 (Grey/Opposite) determines the room color on Tuesday. If Coin 2 is "Opposite," the Tuesday room is the opposite color of Monday's; if it is "Grey," the Tuesday room is always grey.

The White-Black-Grey (WBG) variant setup across four possible states: WG, WO, BO, and BG.

In this model, Beauty is awoken twice. The information she receives (the room color) is distinguishable, yet the setup remains symmetric. This symmetry is the key to the Dutch book.

The Dutch Book Construction

The proposed Dutch book consists of two distinct bets:

  1. Bet 1: Offered on Sunday (before the experiment), costing 20 and paying 42 if Coin 2 is Grey.
  2. Bet 2: Offered only when Beauty finds herself in a White or Black room, costing 24 and paying 33 if Coin 2 is Opposite.

An EDT agent, who views their current action as evidence for what their "other selves" would do in symmetric situations, is led to accept both bets. Because the agent in a White room assumes their "Black room self" would make the same choice, they coordinate into a losing strategy.

The Dutch book for the WBG variant, demonstrating a guaranteed net loss of 2 units across all possible outcomes.

As shown in the table above, accepting all bets results in a total gain of -2 regardless of the coin toss outcomes (WG, WO, BO, or BG). This represents a mathematical guarantee of loss.

Analysis: Why EDT Fails

The failure of EDT in this scenario stems from how agents calculate expected utility. When Beauty is in a White room, she only considers worlds where a White room awakening is possible. This selection bias filters out certain losing scenarios, making the bet appear favorable in isolation.

Crucially, Conitzer’s WBG Dutch book works against EDT agents regardless of whether they are Halfers or Thirders. It demonstrates that the "safety" of EDT in the standard Sleeping Beauty problem was a result of the decision points being identical (indistinguishable). Once the information is distinguishable but symmetric, the coordination logic of EDT leads to irrationality.

Conclusion

The WBG variant suggests a fundamental vulnerability in Evidential Decision Theory regarding self-locating beliefs. While Causal Decision Theory (CDT) agents avoid this trap by evaluating each decision in isolation, EDT agents are lured into coordinated losses. This suggests that for a decision theory to be robust against Dutch books in multi-awakening scenarios, it must account for symmetric information without over-relying on evidential coordination.

Summary updated

Abstract

In the context of the Sleeping Beauty problem, it has been argued that so-called “halfers” can avoid Dutch book arguments by adopting evidential decision theory. I introduce a Dutch book for a variant of the Sleeping Beauty problem and argue that evidential decision theorists fall prey to it, whether they are halfers or thirders. The argument crucially requires that an action can provide evidence for what the agent would do not only at other decision points where she has exactly the same information, but also at decision points where she has different but “symmetric” information.
BibTeX
@inproceedings{conitzer2015dutch,
  author = {Vincent Conitzer},
  title = {{A Dutch book against sleeping beauties who are evidential decision theorists}},
  booktitle = {Synth},
  year = {2015},
  url = {https://openreview.net/forum?id=qXSpCzyd6l},
}

Loading related content...

Figure