Strategy Masking: A Method for Guardrails in Value-based Reinforcement Learning Agents

The use of reward functions to structure AI learning and decision making is core to the current reinforcement learning paradigm; however, without careful design of reward functions, agents can learn to solve problems in ways that may be considered ``undesirable" or ``unethical. Without thorough understanding of the incentives a reward function creates, it can be difficult to impose principled yet general control mechanisms over its behavior. In this paper, we study methods for constructing guardrails for AI agents that use reward functions to learn decision making. We introduce a novel approach, which we call strategy masking, to explicitly learn and then suppress undesirable AI agent behavior. We apply our method to study lying in AI agents and show that strategy masking can effectively modify agent behavior by suppressing, or actively penalizing, the reward dimension for lying such that agents act more honestly while not compromising their ability to perform effectively.

Prediction Instability in Machine Learning Ensembles

In machine learning ensembles predictions from multiple models are aggregated. Despite widespread use and strong performance of ensembles in applied problems little is known about the mathematical properties of aggregating models and associated consequences for safe, explainable use of such models. In this paper we prove a theorem that shows that any ensemble will exhibit at least one of the following forms of prediction instability. It will either ignore agreement among all underlying models, change its mind when none of the underlying models have done so, or be manipulable through inclusion or exclusion of options it would never actually predict. As a consequence, ensemble aggregation procedures will always need to balance the benefits of information use against the risk of these prediction instabilities. This analysis also sheds light on what specific forms of prediction instability to expect from particular ensemble algorithms; for example popular tree ensembles like random forest, or xgboost will violate basic, intuitive monotonicity and fairness properties.