Non-maximizing policies that fulfill multi-criterion aspirations in expectation
Simon Dima, Simon Fischer, Jobst Heitzig, Joss Oliver
TL;DR
The paper tackles the challenge of sequential decision-making under multi-criteria goals by eschewing scalar reward maximization in favor of aspiration-based planning. It presents a framework where the agent maintains and propagates convex aspiration sets over the evaluation space, using simplex-based approximations of feasibility sets and tracing maps to guarantee aspiration satisfaction within a fixed computational budget. The method yields non-maximizing policies that preserve flexibility for safety criteria and can be augmented with information, performance, and safety considerations via a softmin action-selection scheme. The work connects to existing MORL and satisficing ideas while offering verifiable guarantees and potential computational advantages in planning and AI safety contexts. It also points to future directions for learning reference policies via RL and extending the approach to broader environments.
Abstract
In dynamic programming and reinforcement learning, the policy for the sequential decision making of an agent in a stochastic environment is usually determined by expressing the goal as a scalar reward function and seeking a policy that maximizes the expected total reward. However, many goals that humans care about naturally concern multiple aspects of the world, and it may not be obvious how to condense those into a single reward function. Furthermore, maximization suffers from specification gaming, where the obtained policy achieves a high expected total reward in an unintended way, often taking extreme or nonsensical actions. Here we consider finite acyclic Markov Decision Processes with multiple distinct evaluation metrics, which do not necessarily represent quantities that the user wants to be maximized. We assume the task of the agent is to ensure that the vector of expected totals of the evaluation metrics falls into some given convex set, called the aspiration set. Our algorithm guarantees that this task is fulfilled by using simplices to approximate feasibility sets and propagate aspirations forward while ensuring they remain feasible. It has complexity linear in the number of possible state-action-successor triples and polynomial in the number of evaluation metrics. Moreover, the explicitly non-maximizing nature of the chosen policy and goals yields additional degrees of freedom, which can be used to apply heuristic safety criteria to the choice of actions. We discuss several such safety criteria that aim to steer the agent towards more conservative behavior.