Search Papers | Poster Sessions | All Posters
Poster B45 in Poster Session B - Thursday, August 8, 2024, 1:30 – 3:30 pm, Johnson Ice Rink
Building Deterministic Causal Graphs Using Reinforcement Learning in Cognitive Tasks with Evolutionary Bias
Ines Aitsahalia1 (), Adithya Gungi1, Pradyumna Sepúlveda1, Kiyohito Iigaya1; 1Columbia University
Despite the importance of uncovering causal, rather than correlational, structures in the world to survival, algorithms for this type of causal learning remain computationally taxing. Recent neural evidence has challenged the ability of reinforcement learning (RL) algorithms to provide a useful approximation. Here, we present a new reinforcement learning model that uses modified successor representations and incorporates evolutionary death avoidance, capturing a wide variety of human structure learning and animal conditioning. To formally capture the risk of learning in the wild we implement a constraint where punishment distributions are inherently heavy-tailed to account for the risk of death. This places the intrinsic value on having a deterministic graph in this framework, parsimoniously capturing a wide range of instrumental and non-instrumental behaviors.
Keywords: reinforcement learning causal inference structure learning risk avoidance