Using ensemble method to gain variance estimation for t-RNN.

(a) alpha parameter estimation, (b) beta parameter estimation. The top panel represents recovery of stationary RL parameters, the middle depicts recovery of abruptly changing RL parameters, and the bottom shows recovery of gradually changing RL parameters. Ground truth (red), stationary Q-learning (green), Bayesian (yellow), our t-RNN with static training set (pink), and t-RNN (blue). Error bars, Bayesian model (in yellow; calculated using the variance over the particles). Error bars, ours (in blue; calculated using 10 different runs of our model).

(PDF)