論文「Blazing the trails before beating the path: Sample-efficient Monte-Carlo planning」紹介スライドです。NIPS2016読み会@PFN(2017/1/19) https://connpass.com/event/47580/ にて。
UiPath Community: Communication Mining from Zero to Hero
Introduction of "TrailBlazer" algorithm
1. BLAZING THE TRAILS BEFORE
BEATING THE PATH:
SAMPLE-EFFICIENT MONTE-
CARLO PLANNING
KATSUKI OHTO
@NIPS2016-YOMI
2017/1/19
2. INTRODUCED PAPER
• Blazing the trails before beating the path:
Sample - efficient Monte-Carlo planning
(JB. Grill, M. Valko and R. Munos)
• NIPS 2016 accepted paper (poster session)
• Abstract starts with “You are a robot…”
• http://papers.nips.cc/paper/6253-blazing-the-trails-before-
beating-the-path-sample-efficient-monte-carlo-planning
3. TRAILBLAZER
• Nested-fashion Monte-Carlo Planning Algorithm
• Problem settings:
MDP (contains MAX nodes and AVG nodes)
Actions per each state : Finite
State transition candidates : Finite or Infinite
• Strong theoretical guarantee
MAX
AVG
4. AIM
• Input : an MDP (Markov Decision Process)
(discount factor 𝛾, maximum number of valid actions 𝐾),
𝜀 (> 0), 𝛿 (0 < 𝛿 < 1)
• Output : estimated value 𝜇 𝜀,𝛿 of current state 𝑠0
• Aim : Get good estimation of real value 𝒱[𝑠0] of current state
such as
ℙ 𝜇 𝜀,𝛿 − 𝒱 𝑠0 > 𝜀 ≤ 𝛿
( ℙ ∙ means probability of ∙ )
with the minimum number of calls to the generative model (state transition function)
5. 1 PLAYER TREE MODEL
IN STOCHASTIC ENVIRONMENT
• Each MAX node means an
opportunity to decide action
• Each AVG node means
stochastic state transition
MAX
AVG
6. ALGORITHM OVERVIEW
• Global Initialization
set 𝜂, 𝜆 as global value
set 𝑚 as an argument of
root node
• Recursive algorithm
log(𝜂/𝛾)
7. ALGORITHM OVERVIEW 2
• In both MAX nodes and AVG nodes,
arguments are
𝑚 (desired branching factor)
and
𝜀 (admissible estimation error)
• If 𝑚 is large, we can search many children, but we need much time
(dilemma)
• If 𝜀 is small, we can search deeply, but we need much time (dilemma)
8. ALGORITHM
FOR AVG NODES
• Input : 𝑚 and 𝜀
• Output : estimated value
• If admissible error 𝜀 is large, ignore
successive reward
• Fill 𝑚 transition samples
(and store immediate reward)
• search all of 𝑚 sampled next states
• return averaged immediate reward +
estimated successive reward
9. ALGORITHM
FOR MAX NODES
• Input : 𝑚 and 𝜀
• Output : estimated value
• Fill candidate action pool ℒ by all valid actions
• U is a value like standard error of estimation
• Search candidate actions repeatedly until
“Only 1 action left” or “Error might be small”
• If “Error might be small”
then return estimated value of best action
else
search best action 1 more time carefully
10. SAMPLE COMPLEXITY OF TRAILBLAER
• Sample Complexity is a measure of performance of algorithm
• If N (the number of next states) is finite,
(
1
𝜀
)
max(2,
log 𝑁𝜅
log
1
𝛾
+𝑜 1 )
on condition that 𝜅 ∈ 1, 𝐾 (in detail in
the paper)
else
(
1
𝜀
)2+𝑑
on condition that 𝑑 is a measure of difficulty to identify near-
optimal nodes