Orseau and Ring, as well as Dewey, have recently described problems, including self-delusion, with the behavior of agents using various definitions of utility functions. An agent's utility function is defined in terms of the agent's history of interactions with its environment. This paper argues, via two examples, that the behavior problems can be avoided by formulating the utility function in two steps: 1) inferring a model of the environment from interactions, and 2) computing utility as a function of the environment model. Basing a utility function on a model that the agent must learn implies that the utility function must initially be expressed in terms of specifications to be matched to structures in the learned model. These specifications constitute prior assumptions about the environment so this approach will not work with arbitrary environments. But the approach should work for agents designed by humans to act in the physical world. The paper also addresses the issue of self-modifying agents and shows that if provided with the possibility to modify their utility functions agents will not choose to do so, under some usual assumptions.

Majha, Arushi, Sarkar, Sayan, Zagami, Davide

$\textit{Embedded agents}$ are not explicitly separated from their environment, lacking clear I/O channels. Such agents can reason about and modify their internal parts, which they are incentivized to shortcut or $\textit{wirehead}$ in order to achieve the maximal reward. In this paper, we provide a taxonomy of ways by which wireheading can occur, followed by a definition of wirehead-vulnerable agents. Starting from the fully dualistic universal agent AIXI, we introduce a spectrum of partially embedded agents and identify wireheading opportunities that such agents can exploit, experimentally demonstrating the results with the GRL simulation platform AIXIjs. We contextualize wireheading in the broader class of all misalignment problems - where the goals of the agent conflict with the goals of the human designer - and conjecture that the only other possible type of misalignment is specification gaming. Motivated by this taxonomy, we define wirehead-vulnerable agents as embedded agents that choose to behave differently from fully dualistic agents lacking access to their internal parts.

Desai, Nishant, Critch, Andrew, Russell, Stuart J.

It is commonly believed that an agent making decisions on behalf of two or more principals who have different utility functions should adopt a Pareto optimal policy, i.e. a policy that cannot be improved upon for one principal without making sacrifices for another. Harsanyi's theorem shows that when the principals have a common prior on the outcome distributions of all policies, a Pareto optimal policy for the agent is one that maximizes a fixed, weighted linear combination of the principals' utilities. In this paper, we derive a more precise generalization for the sequential decision setting in the case of principals with different priors on the dynamics of the environment. We refer to this generalization as the Negotiable Reinforcement Learning (NRL) framework. In this more general case, the relative weight given to each principal's utility should evolve over time according to how well the agent's observations conform with that principal's prior. To gain insight into the dynamics of this new framework, we implement a simple NRL agent and empirically examine its behavior in a simple environment.

Desai, Nishant, Critch, Andrew, Russell, Stuart J.

Ortega, Pedro A., Braun, Daniel A.

Rewards typically express desirabilities or preferences over a set of alternatives. Here we propose that rewards can be defined for any probability distribution based on three desiderata, namely that rewards should be real-valued, additive and order-preserving, where the latter implies that more probable events should also be more desirable. Our main result states that rewards are then uniquely determined by the negative information content. To analyze stochastic processes, we define the utility of a realization as its reward rate. Under this interpretation, we show that the expected utility of a stochastic process is its negative entropy rate. Furthermore, we apply our results to analyze agent-environment interactions. We show that the expected utility that will actually be achieved by the agent is given by the negative cross-entropy from the input-output (I/O) distribution of the coupled interaction system and the agent's I/O distribution. Thus, our results allow for an information-theoretic interpretation of the notion of utility and the characterization of agent-environment interactions in terms of entropy dynamics.