Summary on "Probabilistic Planning with Sequential Monte Carlo Methods" (ICLR2019)

• probabilistic inference over optimal trajectories 
• planning in continuous domains
• fixed computational budget
• An agent should able to predict the consequences of actions and explain how they will react. 
• Use a model to plan future actions. Some examples for models to plan future actions are Monte Carlo Tree Search (MCTS), cross entropy methods (CEM) and iterative linear quadratic regulator (iLQR)]
• MCTS: discrete games and known dynamics
• CEM: assume the distribution over future optimal trajectories to be Gaussain. 
• iLQR: assume the dynamics are locally linear-Gaussian.