强化学习2023版第一讲 德梅萃·P. 博赛卡斯（Dimitri P. Bert

On-Line Play algorithm.

Online tree search.

So, search all the moves and determine the final values. Determine the move based on the final values.

以果决行。

Off-Line Training in AlphaZero: Approximation Policy Iteration (PI)

a value neural net through training

a policy neural net through training

on-line player plays better than the off-line-trained player.

Central role of Newton's method?

mathematical connection?

跳了这部分。

Reference page.

Terminology.

RL uses Max/Value

DP uses Min/Cost

• Reward of a stage = (Opposite of ) cost of a stage
• State value = (Opposite of) State cost
• Value (or state-value) function = opposite of Cost function

Controlled system terminology

• Agent = Decision maker or controller
• Action = Decision or control
• Environment = Dynamic system

Methods terminology

• Learning = Solving a DP-related problem using simulation
• Self-learning (or self-play in the context of games) = Solving a DP problem using simulation-based policy iteration.
• Planning v.s. Learning distinction = Solving a DP problem with model based v.s. model-free simulation

Notations.

two types: transition probability/discrete-time system equation.

Finite Horizon Deterministic Optimal Control Model

a system ends at stage x_N.

A Special Case: Finite Number of States and Controls.

主要就是说也是shortest path...

Principle of Optimality:

THE TAIL OF AN OPTIMAL SEQUENCE IS OPTIMAL FOR THE TAIL SUBPROBLEM.

If there exists a better solution for the tail subproblem, then we will take that part instead of the current one. Hence, the principle of optimality holds.

From One Tail Subproblem to the Next.

I think for this part, it is to tell us that we can use backward method to solve the problem...

DP Algorithm: Solves all tail subproblems efficiently by using the Principle of Optimality.

中间讲了两个例了跳了。

General Discrete Optimization.

Connect DP to Reinforcement Learning..

Use approximation J^\tilda s instead of J^\star s. (off-line training)

Generate all the approximations.

Then, going forward, to find u^\tilda_k (on-line play)

Extentions:

Stochastic finite horizon problems: x_{k+1} is random

Infinite horizon problems: instead of ending at stage N...

Stochastic partial state information problems:

do not know the state information perfectly

MINIMAX/game problems

课程要求~跳啦