We humans have potential to self-actualize far beyond our immediate or even our own objectives, and our overall life experience is an open-book of this endeavour. However, such sacrifice is not as common in the world of deep reinforcement learning agents. Standard reinforcement learning agents practically embody the behaviorist fear-driven, classically-conditioned interpretation of behavior. Survival-oriented emotions like fear are certainly useful in momentarily narrowing focus to motivate brief behavioral responses. However, the spectrum of human affective experience is complemented by positive emotions such as curiosity, joy, and love, which though rarely directly serving any explicit objective, actually broaden and build on social, cognitive, and behavioral skills for many potential needs. I apply the Broaden and Build theory of positive emotions to the design and training of a novel deep reinforcement learning architecture Affective RL (AffectR) situated in a nonstationary lifelong, multimodal, multiagent, mixed-mode setting. The population is placed in progressively more human-like environments culminating in a photorealistic simulator where interventions are made when necessary to help each AffectR agent realize its full potential. Code: tinyurl.com/affectr
This poster was presented at the 2021 UTA-sponsored Broaden and Build Conference in Arlington, Texas:
Post script: I regret to write that I never completed work on this project after the conference. So although it is written in the perfect tense, it is not a completed project. I hope to incorporate the ideas into future work though.