Pomrn

Pomrn is a term used in some research contexts to denote a class of decision problems that combine partial observability with a stochastic reward structure. The exact meaning of the acronym is not standardized, and its usage can vary between authors. In many discussions, pomrn is related to, or derived from, partially observable decision models such as POMDPs (partially observable Markov decision processes) and can be interpreted as either a partially observable Markov reward process or a partially observable decision framework with a focus on rewards.

In broad terms, a pomrn involves a hidden state that evolves over time, an agent that receives noisy or incomplete observations, and a reward signal that depends on the latent state and possibly the chosen action. Unlike a fully observable MDP, the agent does not directly observe the true state and must maintain a belief distribution over possible states based on history of actions and observations. The objective is typically to select actions to maximize expected cumulative or discounted rewards over a horizon.

Common solution approaches mirror those used for POMDPs, including belief state tracking via Bayesian filtering, point-based or approximate value iteration, and policy search methods that operate on histories or belief representations, including recurrent neural networks. Pomrn models have applications in robotics, autonomous systems, and finance, where decision making under partial information and uncertain rewards is essential.

See also: POMDP, MDP, partially observable decision processes, reinforcement learning.