'Indifference' methods for managing agent rewards

12/18/2017
by   Stuart Armstrong, et al.
0

Indifference is a class of methods that are used to control a reward based agent, by, for example, safely changing their reward or policy, or making the agent behave as if a certain outcome could never happen. These methods of control work even if the implications of the agent's reward are otherwise not fully understood. Though they all come out of similar ideas, indifference techniques can be classified as way of achieving one or more of three distinct goals: rewards dependent on certain events (with no motivation for the agent to manipulate the probability of those events), effective disbelief that an event will ever occur, and seamless transition from one behaviour to another. There are five basic methods to achieve these three goals. This paper classifies and analyses these methods on POMDPs (though the methods are highly portable to other agent designs), and establishes their uses, strengths, and limitations. It aims to make the tools of indifference generally accessible and usable to agent designers.

READ FULL TEXT

page 1

page 2

page 3

page 4

research
07/14/2021

Experimental Evidence that Empowerment May Drive Exploration in Sparse-Reward Environments

Reinforcement Learning (RL) is known to be often unsuccessful in environ...
research
11/27/2015

Shaping Proto-Value Functions via Rewards

In this paper, we combine task-dependent reward shaping and task-indepen...
research
11/28/2018

Unsupervised Control Through Non-Parametric Discriminative Rewards

Learning to control an environment without hand-crafted rewards or exper...
research
05/11/2020

Maximizing Information Gain in Partially Observable Environments via Prediction Reward

Information gathering in a partially observable environment can be formu...
research
07/29/2019

Reinforcement with Fading Memories

We study the effect of imperfect memory on decision making in the contex...
research
04/04/2022

Continuously Discovering Novel Strategies via Reward-Switching Policy Optimization

We present Reward-Switching Policy Optimization (RSPO), a paradigm to di...
research
06/15/2020

Pessimism About Unknown Unknowns Inspires Conservatism

If we could define the set of all bad outcomes, we could hard-code an ag...

Please sign up or login with your details

Forgot password? Click here to reset