Please enable JavaScript.
Coggle requires JavaScript to display documents.
intrinsic motivation & RL (intrinsically motivated exploration…
intrinsic motivation & RL
exploration in domain w sparse reward - use info-theoretic quantities to describe intrinsically motivated behaviour
curiosity-driven
in discrete state-action space
Stadie (2015)
in continuous state-action space
Pathak (ICML2017) - quantify curiosity as prediction error (minimising) - dynamic-based curiosity w embedding observation
Houthooft (NIPS2016) - quantify curiosity as information gain (maximising)
Savinov (2018) - curiosity measured by reachability
visitation count - count-based exploration
in discrete state-action space
Bellemare (NIPS2016) - count-based exploration using pseudo-count derived w density models & relate to information gain & prediction gain
Lopes (2012) - assign reward according to estimated learning progress which is computed based on history of transitions and visitation counts
computational models of IM
(closely related to RL) - J. Schmidhuber - quantify fun, curiosity as information gain measured as KL divergence
(robotics) - PY Oudeyer - quantify intrinsic motivation as learning progress
(RL) - A Barto
reward shaping
Ng (ICML1999) - potential-based reward shaping
Singh (2010) - fitness-based reward shaping
psychological IM
Deci & Ryan - distinction betw intrinsic & extrinsic motivation
Berlyne - factors underlying IM effects: novelty, surprise, incongruity, complexity; moderate novelty has the highest hedonic value
intrinsically motivated exploration
visitation count-based exploration
Bellemare (NIPS2016)
learning progress
Lopes (NIPS2012), Oudeyer
prediction error
Pathak (ICML2017)
information gain
Still-Precup (2012), Houthooft (NIPS2016), Achiam (2018)
posterior sampling exploration