11:12
Autonomy Talks meeting Zachary Sundberg
benefits and curse of POMDP
Uncertainty
aleatory uncertainty
observation does not improve predictive power
Markov Decision Process
Curse of Dimensionality
exponential growth if we discretize state space
epistemic uncertainty
observation does improve predictive power
Reinforcement learning
use neural networks to partially estimate outcomes at higher dimensions, somewhat side-stepping the Curse of Dimensionality
static
unknown properties don’t change over time
dynamic
unknown properties may change over time
POMDP
interaction uncertainty
game theoretic
partially observed Markov Decision Process
Bayesian updating of beliefs
like an Markov Decision Process on the belief space
breaking the curse of POMDP
consider integration
estimation is integration
has Curse of Dimensionality
use Monte Carlo method simulation
partially observable markov game
POMDP with multiple agents
pomdps.jl