在強化學習領域，區分了有限的分節式任務以及無窮情況的連續性任務，並對這兩種任務形式設立了相應的累積獎勵函數。然而，最近的研究發現，使用傳統的累積獎勵公式，實際上會存在跟策略有關的縮放因子，進而導致梯度下降算法的理論偏差值。本文以此為出發點，提供無需重構程式也能解決縮放因子的無窮蒙地卡羅樹框架與無關乎策略的訓練方法。接著強調環境、訓練方式、訓練資料、超參數等人為調整的資料，本質上也是人類在面對「訓練代理的環境」所採取的策略，最終使代理能解決「不存在於環境與資料」的任務。最後，使用無窮蒙地卡羅樹的框架，我發現有限時間的分節式任務和無窮時間的連續性任務本質上是相同的。並且在考量縮放因子後的累積獎勵函數和含折扣率的獎勵函數，其實是更一般的時步權重累積函數的特例。;In the field of reinforcement learning， a distinction is made betweenfinite episodic tasks and infinite continuous tasks， andcorresponding return functions are established for these two types oftasks. However， recent research has found that using the traditionalreturn formula involves a policy-related scaling factor， which leadsto theoretical deviations in the gradient descent algorithm. Startingfrom this point， this paper provides an infinite Monte Carlo Treeframework and corresponding policy-independent training method thatcan solve the scaling factor without refactoring the program. Theemphasis on human-adjusted data such as environment， trainingmethods， training data， and hyperparameters is essentially a policyhumans adopt when facing the "training environment from an agent".Ultimately enabling the agent to solve tasks that "do not exist inthe environment and data." Finally， using the framework of infiniteMonte Carlo Tree， I found that finite time-step episodic tasks andinfinite time-step continuing tasks are essentially the same.Furthermore， the return function after considering the scalingfactor， no matter episodic tasks or continuous tasks， are actuallyspecial cases of the more general case time-step weight returnfunction.
