Algorithms.md

算法

Short Name	Comment 1	Comment 2
STRIPTs
BrFS(BFS)	Breadth-first-search	- Completeness - Optimality(if costs are uniform)
DFS	Deep-first-search	- Not completeness - Not optimality
ID	Iterative Depening	- Completeness - Optimality
GBFS	Greedy Best-First Search
BFWS	Best-First Width Search
$h^*$		- Optimal one (theoretical)
$h^+$		Safe, Admissible (drop deletion)
$h^{add}$		Safe, not admissible (based on $h^+$)
$h^{max}$		Safe, admissible (based on $h^+$)
$IW(k)$	- if $novetly(s) > k$, then prune	- try to solve problem first in $IW(1)$, if not solved, then $IW(2)$, ... - novetly: smallest subset of atoms (which is first showing up) size of the new state
MDP	Or	fully observable, probabilistic state models - value iteration:(update value via last iteration value) - policy iteration:(update policy via existing policy)
MCTs
UCT
Q-Learning	off-policy optimistic unsafe or risky
SARSA	safe
n-step SARSA
Nash Equation
Mix strategies: Indifferency		- think of $A$ and $B$, the indifferency for $A$ is that the probability $X$ of selecting action $a$ that makes $B$ have the same reward.

Reinforcement Learning Cons