Generalizations of the Theory of Linearly Solvable MDPs

Published on 2012-10-163360 Views

Krishnamurthy Dvijotham

We present various generalizations of the theory of linearly solvable optimal control. One generalization is the extension to the game theoretic and risk-sensitive setting, which requires replacing KL

Workshop on Statistical Physics 2012 - Granada

Related categories

Presentation

Generalizations of the Theory of Linearly Solvable MDPs00:00

Table of Contents00:20

Markov Games (MGs)01:12

I Like Linear Equations03:57

Rényi divergences - 104:56

Rényi divergences - 205:42

Linearly Solvable Markov Games (LMGs): Dynamics06:12

LMGs: Solution08:28

Car on a Hill - 112:23

Car on a Hill - 214:00

Crossing a Narrow Bridge - 114:31

Crossing a Narrow Bridge - 214:54

Stabilizing an unstable system - 115:13

Stabilizing an unstable system - 215:29

Outline: Policy Gradients15:59

Deterministic MDPs with Stochastic Policies16:31

Finite Horizon Policy Gradients17:18

Relationship to Other Algorithms - 118:22

Relationship to Other Algorithms - 219:54

Second Order Methods20:03

Conclusions and Future Work - 121:13

Conclusions and Future Work - 221:25

References21:52