Since Newcomb’s problem has been proposed in 1969, there has been disagreement about what the right basis for rational decision-making in Newcomb-like problems should be. The two main competing theories are causal and evidential decision theory. However, various alternatives have since been proposed. This page is an attempt at listing and, to some extent, categorizing all of these decision theories, including obscure, unpublished ones. If you are aware of a decision theory that we have not listed, let us know. We plan to continually update this page.
There are a few properties that are shared by many decision theories. On this page, we will use categories to denote which decision theories have which properties:
- Common cause versus logical correlations – Treutlein and Oesterheld (unpublished) argue that Newcomb-like decision problems can be divided into two classes based on how the correlation between decision and payoff arises. In reference class cases, the correlation is based on the logical similarity between the agent’s decision algorithm and some aspect of the environment. For example, in a prisoner’s dilemma against a psychologically similar opponent, the correlation arises from the fact that if I cooperate, it is more likely that the opponent, who uses a similar decision algorithm, cooperates as well. In common cause cases, the correlation is based on our decision and the payoff having some common cause. For example, in the Smoking Lesion, it may be that a gene causes both our decision to smoke and our having cancer. Treutlein and Oesterheld point out that people commonly find it more intuitive to follow EDT in reference class cases, but more intuitive to follow CDT in common cause cases. Consequently, many of the alternative decision theories that have been proposed incorporate this distinction and take reference class correlations but not common cause correlations into account.
- Causal dominance distinction – In Newcomb’s problem, no matter the state of the environment, two-boxing is better. In other words, two-boxing causally dominates in Newcomb’s problem. (For the distinction between dominance with causal independence and evidential independence, see section 1.6 of Ahmed 2014 or section 5.1 of Joyce 1999.) But causal dominance is not necessary for CDT and EDT to disagree. Thus, there are Newcomb-like problems without causal dominance. These cases are sometimes called Egan cases. An example is the psychopath button (Egan 2007). Some argue that one should follow CDT’s recommendation in cases of causal dominance but not (or not necessarily) in Egan cases (cf. Ahmed 2012) and have consequently proposed decision theories that are sensitive to this difference.
- Updatelessness – See my blog post “Thoughts on Updatelessness“. Note that the “strength” of different decision theories’ updatelessness may vary. I will err on the side of using the term liberally, here. For example, I will call decision theories updatelessness as soon as they don’t need precommitments to win in Parfit’s hitchhiker. However, updateless arising from anthropic uncertainty does not suffice.
List of decision theories
If a category box is empty, then it means that we either forgot to include it or are unsure which is true, e.g. whether a decision theory makes distinctions based on causal dominance or not. Of course, the given category statements are not guaranteed, either, especially in the cases of not fully formally specified decision theories. Sometimes they are merely based on extrapolations from the statements made by the proponents of the decision theory about example decision problems.
The list is very roughly sorted in increasing order of “obscurity” with formally published decision theorists coming first, followed by decision theories that have not been formally published, roughly in order of how much they have been discussed.
|Decision theory||common cause correlations||logical correlations||causal dominance distinction||updateless|
|Causal decision theory (Gibbard and Harper 1981; Joyce 1999; Lewis 1981; Skyrms 1982)||no||no||no||no|
|Evidential decision theory (Ahmed 2014; Almond 2010; Eells 1982; Horgan 1981; Price 1986)||yes||yes||no||no|
|Spohn’s (2003; 2012) variant of CDT||no||yes||no||no|
|Arntzenius’ (2008) deliberational decision theory||yes|
|Wedgwood’s (2013) benchmark theory||yes|
|Dohrn’s (2015) modified EDT||yes|
|Gustafsson’s (2011) ratificationism||yes|
|Price’s (2012) EviCausalist decision theory||no (?)||yes (?)|
|Soares and Levinstein’s (2017) functional decision theory||no (?)||yes||no||yes|
|Poellinger’s (2013) variant of CDT||no||yes|
|Barnett’s (2014) variant of CDT||yes|
|Fisher’s (n.d.) disposition-based decision theory|
|Shpitser’s variant of CDT||no|
|Yudkowsky’s (2010) Timeless decision theory||no (?)||yes||no||yes|
|Dai’s Updateless decision theory
|Ambient decision theory|
|Asymptotic decision theory|
|Drescher’s metacircular decision theory|
|Goertzel’s (2010) counterfactual reprogramming decision theory||yes||yes|
|Proof-based decision theory|
Partial decision theories
In this section, I list decision theories that only attempt to solve part of the problem or are meant to be combined with other decision theories or parts of a decision theory.
- Meta-decision theories, hybrid theories and decision-theoretic uncertainty (Nozick 1993; Ledwig 2000, ch. 4; MacAskill 2016; Weatherson n.d., chapter 18.5) are attempts to combine multiple decision theories based on one’s credence in these theories.
- Armstrong’s (2011) anthropic decision theory attempts to solve the problem of acting in situations of anthropic uncertainty.