Paper Drafts

Here are draft versions of two short papers. They are on machine ethics and decision theory in the context of reinforcement learning. Comments are welcome.

Altruistic Reinforcement Learning

Abstract:

This article presents a modification of reinforcement learning where an agent’s action lead to rewards being received by a second agent interacting with same environment. This model can be useful in the development of powerful AIs. Agent policies are proposed for dealing with observable rewards, with non-observable rewards in perfectly rational agents, and with non-observable rewards in bounded rational agents.

Newcomblike Problems and Optimal Agents

Abstract:

This article discusses the family of Newcomblike problems in the context of reinforcement learning. It reframes the problem of rational decision making as one of obtaining maximal rewards in a wide range of environments. Newcomblike problems are characterized by correlations between agent and environment policies. An optimal policy, taking into account these correlations, is given for known environments. For unknown environments, a quality criterion for policies is formulated.