This work was done during one weekend by research workshop participants and does not represent the work of Apart Research.
Accepted at the 
 research sprint on 
September 25, 2023

Preserving Agency in Reinforcement Learning under Unknown, Evolving and Under-Represented Intentions

This paper investigates several techniques to implement altruistic RL while preserving agency. Using a two-player grid game, we train a helper agent to support a lead agent in achieving their goals. By training the helper without showing them the goal and resampling the goals to rebalance for unequal value distributions, we demonstrate that helpers can act altruistically without observing the goals of the lead. We also initiate exploration of a technique to encourage corrigibility and respect for personal agency by resampling the leads values during training time, and point towards how these techniques could be used to translate into real-world situations through meta-learning.

Harry Powell, Luigi Berducci
4th place
3rd place
2nd place
1st place
 by peer review