uncertainty
- #aleatoric uncertainty is statistical uncertainty, i.e. in the frequentist view noisy measurements.
- #epistemic uncertainty is systematic uncertainty, i.e. model aspects not considered, such as physical constraints like gravity, etc.
Reinforcement learning
In the context of [[reinforcement learning]], according to [[decision-making-book]] uncertainty comprises:
- outcome uncertainty
- what effect will our action have
- model uncertainty
- our model of the problem is uncertain
- state uncertainty
- the true state is uncertain
- can be modelled as a [[partially-observable-markov-decision-process]]
- interaction uncertainty
- behavior of other agents is uncertain
Backlinks
variational autoencoder
A probabilistic [[architecture]] from the family of [[autoencoders]]. The self-learned representation is used to parameterize a probability distribution (i.e. the [[posterior distribution]]), from which a decoder can draw samples from to generate a range of outputs. We can either directly predict the mean of the distribution, or perform sampling over the distribution to obtain #uncertainty estimates.