Abstract
This paper proposes a formal separation between an entity's epistemic system and its decision system to resolve Newcomb-like problems. We demonstrate that Bayesian counterfactuals are estimated by the epistemic system rather than the decision system. Under this framework, we prove that for certain problems, an epistemic system necessarily expects the entity to make a counterfactually suboptimal decision. By applying this to a generalization of Newcomb’s paradox, we show that the optimality of the 1-Box strategy depends on the player's prior regarding the predictor's information. If the predictor is believed to be quasi-omniscient due to a data supset, 1-boxing is counterfactually preferred; otherwise, 2-boxing may prevail. This analysis suggests that understanding complex agents—whether human or artificial—requires a modular view of their internal epistemic, decision, and maintenance systems.
Introduction
Newcomb’s Paradox has long served as a fundamental challenge in decision theory, creating a rift between Causal Decision Theory (CDT) and Evidential Decision Theory (EDT). The problem involves two boxes: Box A (opaque, potentially containing $1,000,000) and Box B (transparent, containing $1,000). A highly reliable predictor, Omega, has already placed the large sum in Box A only if it predicted the player would take Box A alone (1-boxing). If it predicted the player would take both (2-boxing), Box A is empty.
CDT advocates for 2-boxing because the physical contents of the boxes are already determined. EDT advocates for 1-boxing because of the high correlation between 1-boxing and wealth. This paper resolves the conflict by sticking to pure Bayesian probability, avoiding ad-hoc operators by deconstructing the agent into distinct functional components.
The Framework: Deconstructing the Agent
To analyze the paradox, we separate the agent into two distinct systems:
- The Epistemic System (Emma): A pure Bayesian observer that infers the state of the world based on available data.
- The Decision System (Dan): The component responsible for executing the final choice.
Crucially, this model assumes imperfect self-knowledge. Emma does not know with 100% certainty what Dan will decide. This uncertainty is mathematically vital; it allows Emma to perform counterfactual reasoning—asking "What would happen if Dan chose X?"—without conditioning on an event with zero probability. This "logical non-omniscience" may stem from computational irreducibility, where Emma cannot know the result of Dan's internal processing without Dan actually completing the computation.
The Impossibility of Guaranteed Optimality
A significant result of this framework is Theorem 1, which establishes that no decision system can guarantee counterfactual optimization in all environments. For any entity with imperfect self-knowledge, there exist Newcomb-like scenarios where the entity will assign a high probability (at least $1 - 1/n$) to making a decision that its own epistemic system identifies as counterfactually bad. Essentially, the agent’s internal structure can lead it to act against its own perceived best interests.
Resolving Newcomb’s Paradox via Information Asymmetry
The paper addresses Newcomb’s paradox by modeling the predictor (Omega) as a Bayesian entity possessing a "supset" of the data available to the player. The "correct" choice is not a universal constant but is determined by Emma’s beliefs about the extent of Omega's information.
The Threshold for 1-Boxing
The counterfactual preference for 1-boxing is governed by the following inequality: $$\frac{r}{R} \leq \frac{\sigma^2}{p(1-p)}$$ Where:
- $r$ is the small reward ($1,000).
- $R$ is the large reward ($1,000,000).
- $p$ is the prior probability of 1-boxing.
- $\sigma^2$ is the variance of the predictor's guess.
Quasi-Omniscience vs. Uncertainty
If Emma believes Omega is "quasi-omniscient"—meaning Omega’s additional data is expected to significantly reduce the uncertainty regarding Dan’s choice—then 1-boxing becomes the counterfactually superior strategy. In this state, Emma concludes that any world where Dan 2-boxes is almost certainly a world where Omega predicted that action and left Box A empty. Conversely, if the predictor's data is not expected to reduce uncertainty sufficiently, Emma’s epistemic system will counterfactually prefer 2-boxing.
Discussion and Implications
This research suggests that the debate between CDT and EDT can be reconciled by viewing counterfactuals as internal Bayesian estimates. Beyond the paradox, this "entity deconstruction" has implications for AI safety and the design of autonomous agents. By separating an entity into Epistemic, Decision, Data Collection, Reward, and Maintenance systems, we can better understand how agents behave when they are "embedded" within the environments they seek to manipulate. Future work may explore how maintenance systems can correct for internal flaws, such as computational limitations or diverging priors between the agent and its environment.
Conclusion
By treating counterfactuals as purely Bayesian inferences made by an internal epistemic system, we provide a mathematical resolution to Newcomb's Paradox. The optimality of a strategy is shown to be a function of the agent's internal model of the predictor's knowledge, offering a robust framework for analyzing agency in both biological and algorithmic systems.
Summary updated