2.845

2023影响因子

(CJCR)

  • 中文核心
  • EI
  • 中国科技核心
  • Scopus
  • CSCD
  • 英国科学文摘

留言板

尊敬的读者、作者、审稿人, 关于本刊的投稿、审稿、编辑和出版的任何问题, 您可以本页添加留言。我们将尽快给您答复。谢谢您的支持!

姓名
邮箱
手机号码
标题
留言内容
验证码

阿法狗围棋系统的简要分析

田渊栋

徐宇锋. 广义分数阶混沌系统的动力学行为. 自动化学报, 2017, 43(9): 1619-1624. doi: 10.16383/j.aas.2017.e150118
引用本文: 田渊栋. 阿法狗围棋系统的简要分析. 自动化学报, 2016, 42(5): 671-675. doi: 10.16383/j.aas.2016.y000001
Xu Yufeng. Dynamic Behaviors of Generalized Fractional Chaotic Systems. ACTA AUTOMATICA SINICA, 2017, 43(9): 1619-1624. doi: 10.16383/j.aas.2017.e150118
Citation: TIAN Yuan-Dong. A Simple Analysis of AlphaGo. ACTA AUTOMATICA SINICA, 2016, 42(5): 671-675. doi: 10.16383/j.aas.2016.y000001

阿法狗围棋系统的简要分析

doi: 10.16383/j.aas.2016.y000001
详细信息
    作者简介:

    田渊栋 脸书人工智能研究所研究员.主要研究方向为深度学习及计算机视觉.2013至2014年曾任谷歌无人车组研究员/软件工程师.2008年毕业于上海交通大学获硕士学位,2013年于美国卡耐基梅隆大学机器人系获博士学位,曾获2013年国际计算机视觉会议(ICCV)马尔奖提名.E-mail:yuandong@fb.com

A Simple Analysis of AlphaGo

More Information
    Author Bio:

    Research scientist in Facebook AI Research, working on deep learning and computer vision. Prior to that, he was a researcher/software engineer in Google Self-driving Car Team in 2013»2014. He received Ph. D. in Robotics Institute, Carnegie Mellon University in 2013, Bachelor and Master degrees in computer science in Shanghai Jiao Tong University. He is the recipient of 2013 ICCV Marr Prize Honorable Mentions.

  • 摘要: 谷歌的围棋系统阿法狗(AlphaGo)在三月的比赛中以4:1的成绩击败了围棋世界冠军李世石, 大大超过了许多人对计算机围棋程序何时能赶上人类职业高手的预期(约10~30年).本文在技术层面分析了阿法狗系统的组成部分, 并基于它过去的公开对局预测了它可能的弱点.
  • Fractional calculus and fractional differential equations have received considerable interest in the recent forty years. Fractional derivative means that the order of differentiation can be an arbitrary real number and even it can be a complex number. Fractional derivative modelling has been applied to many scientific and engineering fields, such as quantum mechanics [1], viscoelasticity and rheology [2], electrical engineering [3], electrochemistry [4], biology [5], biophysics and bioengineering [6], signal and image processing [7], mechatronics [8], and control theory [9]-[11]. Although few mathematical issues of fractional derivative remain unsolved, most of the difficulties have been overcome, and the applications of fractional calculus in above fields indicate that the fractional models can depict the property and behavior of a real-world problem more accurately. For a comprehensive review of fractional calculus, we refer readers to some monographs [12]-[14] and references therein. In contrast to integer order derivative, the way of identifying fractional derivative is not unique. There are several types of definitions, such as Riemann-Liouville derivative, Caputo derivative, Grünwald-Letnikov derivative, and so on. More details can be found in [13, Chapter 2]. In the recent years, the study of dynamical system with fractional order derivative becomes more and more popular [15]-[19]. Moreover, the dynamics in fractional dynamical system is more interesting.

    Returning back to the fractional derivative, since it has several different definitions, how to develop a generalized form which can unify all the existing fractional derivatives becomes one important topic in fractional calculus [20]-[22]. Recently, a class of new generalized fractional integral and generalized fractional derivative is introduced in [22]. The new generalized fractional integral and generalized fractional derivative depend on a scale function and a weight function, which makes them more general. When the scale function and the weight function reduce to some specific cases, the generalized fractional operators will reduce to Riemann-Liouville fractional integral, Riemann-Liouville fractional derivative and Caputo fractional derivative and so on. However, the study of this new generalized fractional integral and generalized fractional derivative are in the very beginning stage now [23]-[26]. In [24], we show that in generalized fractional diffusion equation, the scale function allows the response domain to be scaled differently. It is required that the scale function should be strictly monotonically increasing or decreasing. A convex increasing scale function will compress the response domain towards to the initial time. A concave increasing scale function will stretch the response domain away from the initial time. The weight function allows the response to be assessed differently at different time, since in many applications, we may require an event to be weighed differently at different time point. For example, modeling of memory of a child may require a heavy weight at current time point, whereas the same for an older person may require more weight on the past. To be an initial attempt of application to chaotic dynamical systems, in this paper, we define a class of new generalized fractional chaotic systems by replacing the original derivatives with the new generalized fractional derivative, then apply a finite difference scheme to study the numerical solutions of two different generalized fractional chaotic systems, namely generalized fractional Lotka-Volterra system (GFLVS) and generalized fractional Lorenz system (GFLS). Their complex dynamics will be discussed, and the dynamic behavior depending on the weight and scale function will be shown graphically.

    The rest of this paper is organized as follows: In Section 2, the preliminaries of fractional calculus are given. The new generalized fractional integral and generalized fractional derivative are shown. A finite difference approach for solving equations with generalized fractional derivative is carried out. In Section 3, we define the chaotic systems using the generalized fractional derivative of Caputo type, i.e., the GFLVS and GFLS. Some interesting dynamics of those two systems are shown graphically. Finally, the conclusions are drawn in Section 4.

    In this section, we introduce the preliminaries of generalized fractional derivatives, and show a proper numerical method for differential equations with such derivatives.

    Let us begin with the common fractional operators. In calculus, the $n$ -fold integral of an integrable function $u(t)$ is defined as

    $ I^{n}u(t)=\overbrace{\int^t_0\cdots\int^t_0}^{n\ {\rm times}}u(s)ds\cdots{ds}= \int^t_0\frac{(t-s)^{n-1}}{(n-1)!}u(s)ds $

    where $t\geq{0}$ , and $u(0)$ is well-defined. Replacing the positive integer $n$ by a real number $\alpha>0$ , we have the following definition.

    Definition 1[13]: The left Riemann-Liouville fractional integral of order $\alpha>0$ of a function $u(t)$ is defined as

    $ \begin{align} \left(I^{\alpha}_{0+}u\right)(t) = \frac{1}{\Gamma(\alpha)}\int^t_0(t-s)^{\alpha-1}u(s)ds \end{align} $

    (1)

    provided the integral is finite, where $\Gamma(\alpha)$ is the Gamma function.

    The Riemann-Liouville fractional integral plays an important role in defining fractional derivatives. There are two basic approaches to define the fractional derivative, i.e., "first integration then differentiation" and "first differentiation then integration". The corresponding fractional derivatives are called Riemann-Liouville fractional derivative and Caputo fractional derivative, and the definitions are given as follows.

    Definition 2[13]:The left Riemann-Liouville fractional derivative of order $n-1 < \alpha < n$ of a function $u(t)$ is defined as

    $ \begin{align} \left(D^{\alpha}_{0+}u\right)(t) = \frac{1}{\Gamma(n-\alpha)}\left(\frac{d^n}{dt^n}\right) \int^t_0(t-s)^{n-\alpha-1}u(s)ds \end{align} $

    (2)

    provided the right side of the identity is finite.

    Definition 3[13]: The left Caputo fractional derivative of order $n-1 < \alpha < n$ of a function $u(t)$ is defined as

    $ \begin{align} \left({^cD}^{\alpha}_{0+}u\right)(t) = \frac{1}{\Gamma(n-\alpha)}\int^t_0(t-s)^{n-\alpha-1}u^{(n)}(s)ds \end{align} $

    (3)

    provided the right side of the identity is finite.

    Besides above, there also exist right Riemann-Liouville integral and derivative, and right Caputo fractional derivative [13]. Mathematically, the Riemann-Liouville and Caputo fractional operators are used in applications frequently. In most real-world models, we always employ the left Caputo fractional derivative. One reason is that we will study generalized fractional dynamical system later, and the derivative is taken with respect to time variable. In physical models, time is always running forward. The other reason is that in the differential equations with Caputo fractional derivative, the initial conditions are taken in the same form as for integer-order differential equations which have clear physical meanings in the practical application and can be easily measured [14]. In what follows, we will introduce the generalized fractional integral and derivative proposed in [22]. They extend nearly all the existing fractional operators. Now we list the generalized fractional integral and derivative defined on positive half axis. They will be used to define the generalized fractional chaotic systems in next section.

    Definition 4[22]: The left generalized fractional integral of order $\alpha>0$ of a function $u(t)$ with respect to a scale function $\sigma(t)$ and a weight function $w(t)$ is defined as

    $ \begin{align} \left(I^{\alpha}_{0+;[\sigma, w]}u\right)(t) = \frac{[w(t)]^{-1}}{\Gamma(\alpha)}\int^{t}_{0} \frac{w(s)\sigma'(s)u(s)}{[\sigma(t)-\sigma(s)]^{1-\alpha}}ds \end{align} $

    (4)

    provided the integral exists, where $\sigma'(s)$ indicates the first derivative of the scale function $\sigma$ .

    Definition 5[22]: The left generalized derivative of order $m$ of a function $u(t)$ with respect to a scale function $\sigma(t)$ and a weight function $w(t)$ is defined as

    $ \begin{align} \left(D^m_{[\sigma, w;L]}u\right)(t) = [w(t)]^{-1}\left[\left(\frac{1}{\sigma'(t)}D_t\right)^m(w(t)u(t))\right] \end{align} $

    (5)

    provided the right-side of equation is finite, where $m$ is a positive integer.

    Definition 6[22]: The Caputo type left generalized fractional derivative of order $\alpha>0$ of a function $u(t)$ with respect to a scale function $\sigma(t)$ and a weight function $w(t)$ is defined as

    $ \begin{align} \left(D^{\alpha}_{0+;[\sigma, w]}u\right)(t) = \left(I^{m-\alpha}_{0+;[\sigma, w]}D^m_{[\sigma, w;L]}u\right)(t) \end{align} $

    (6)

    provided the right-side of equation is finite, where $m-1\leq$ $\alpha$ $ < $ $m$ , and $m$ is a positive integer. Particularly, when $0 < $ $\alpha$ $ < $ $1$ , we have

    $ \begin{align} \left(D^{\alpha}_{0+;[\sigma, w]}u\right)(t) = \frac{[w(t)]^{-1}}{\Gamma(1-\alpha)}\int^{t}_{0} \frac{[w(s)u(s)]'}{[\sigma(t)-\sigma(s)]^{\alpha}}ds. \end{align} $

    (7)

    Now we introduce a finite difference method for solving differential equations with generalized fractional derivative. Consider the following generalized fractional differential equation:

    $ \begin{align} \begin{cases} \left(D^{\alpha}_{0+;[\sigma, w]}u\right)(t)=f(t, u(t)), \quad 0 < t\leq T\\ u(0)=u_0 \end{cases} \end{align} $

    (8)

    where $0 < \alpha < 1$ and $T$ is the final time. Without loss of generality, on a uniform mesh $0=t_0 < t_1 < \cdots < $ $t_j < $ $t_{j+1} < \cdots < t_N=T$ , the Caputo type generalized fractional derivative of $u(t)$ can be approximated as

    $ \begin{align} (D^{\alpha}_{0+;[\sigma, w]}& u)(t_{j+1}) \nonumber\\ &= \frac{[w(t_{j+1})]^{-1}}{\Gamma(1-\alpha)}\int^{t_{j+1}}_{0}\frac{[w(s)u(s)]'} {[\sigma(t_{{j+1}})-\sigma(s)]^{\alpha}}ds \nonumber\\ &= \frac{w_{j+1}^{-1}}{\Gamma(1-\alpha)}\sum^{j}_{k=0}\int^{t_{k+1}}_{t_k} \frac{[w(s)u(s)]'}{\left[\sigma(t_{{j+1}})-\sigma(s) \right]^{\alpha}}ds \nonumber\\ & \approx \frac{w_{j+1}^{-1}}{\Gamma(1-\alpha)}\sum^{j}_{k=0}\int^{t_{k+1}}_{t_k} \frac{\frac{w_{k+1}u_{k+1}-w_ku_k}{t_{k+1}-t_k}} {\left[\sigma_{j+1}-\sigma(s)\right]^{\alpha}}ds \nonumber\\ & \approx\sum^{j}_{k=0}\left(A^j_ku_{k+1}-B^j_ku_k\right) \end{align} $

    (9)

    where

    $ \begin{align*} A^j_k =&\ \frac{w^{-1}_{j+1}w_{k+1}}{\Gamma(2-\alpha)(\sigma_{k+1}-\sigma_k)} \\ & \times \left[(\sigma_{j+1}-\sigma_k)^{1-\alpha}- (\sigma_{j+1}-\sigma_{k+1})^{1-\alpha}\right]\\ B^j_k =&\ \frac{w^{-1}_{j+1}w_{k}}{\Gamma(2-\alpha)(\sigma_{k+1}-\sigma_k)} \\ & \times \left[(\sigma_{j+1}-\sigma_k)^{1-\alpha}- (\sigma_{j+1}-\sigma_{k+1})^{1-\alpha}\right] \end{align*} $

    $k=0, 1, 2, \ldots, j$ , $u_j=u(t_j)$ , $w_j=w(t_j)$ , and $\sigma_j=\sigma(t_j)$ .

    Therefore, we obtain the finite difference scheme:

    $ \begin{align} \sum^{j}_{k=0}\left(A^j_ku_{k+1}-B^j_ku_k\right)=f(t_{j+1}, u_{j+1}) \end{align} $

    (10)

    and the corresponding iteration scheme as

    $ \begin{align} u_{j+1}=\begin{cases} \frac{1}{A^j_j}\left[f_j-\sum\limits^{j-1}_{k=0} \left(A^j_ku_{k+1}-B^j_ku_k\right)+B^j_ju_j \right], \\ \qquad \qquad \qquad \qquad \qquad\qquad j=1, 2, \ldots, N-1\\ \frac{1}{A^0_0}\left(f_0+B^0_0u_0\right), \qquad\qquad \ \, j=0 \end{cases} \end{align} $

    (11)

    where $f_j=f(t_j, u_j)$ .

    In what follows, we will apply this method to solve the generalized fractional chaotic systems. The numerical analysis of the above scheme can be found in [26].

    In this section, we introduce two nonlinear dynamical systems but redefine them with Caputo type generalized fractional derivative. The classical and fractional senses are special cases of the new generalized fractional system below.

    Replacing the derivative with the generalized fractional derivative defined by (7), we define the generalized fractional Lotka-Volterra system (GFLVS) as

    $ \begin{align} \begin{cases} D^{\alpha_1}_{0+;[\sigma, w]}x = ax - bxy + mx^2 - sx^2z\\ D^{\alpha_2}_{0+;[\sigma, w]}y = -cy +dxy\\ D^{\alpha_3}_{0+;[\sigma, w]}z = -pz + sx^2z \end{cases} \end{align} $

    (12)

    where $0 < \alpha_1, \alpha_2, \alpha_3 < 1$ ( $\alpha_1$ , $\alpha_2$ , $\alpha_3$ can be the equal or different) are the orders of the derivative and parameters $a$ , $b$ , $c$ , $d$ are positive. $a$ represents the natural growth rate of the prey in the absence of predators, $b$ represents the effect of predator on the prey, $c$ represents the natural death rate of the predator in the absence of prey, $d$ represents the efficiency and propagation rate of the predator in the presence of prey, and $m$ , $p$ , $s$ are positive constants.

    By selecting the parameters $a=1$ , $b=1$ , $c=1$ , $d=1$ , $m$ $=$ $2$ , $s=2.7$ , $p=3$ and the initial condition $[x_0, y_0, z_0]$ $=$ $[1.5, 1.5, 1.5]$ , when $\alpha_1=\alpha_2=\alpha_3=0.95$ , (12) represents the generalized fractional Lotka-Volterra chaotic system and the phase portraits of the system (12) are described through Figs. 1(a) and 1(b). In Fig. 1(a), the chaotic phenomenon is shown. Moreover, the GFLVS reduces to the fractional Lotka-Volterra system as $\sigma(t)=t$ and $w(t)=1$ . In Fig. 1(b), we see that when the scale function is specified as a power function, and the weight function is taken as an exponential function, the chaotic attractor vanishes and then a stable equilibrium point appears.

    图 1  Phase portraits of GFLVS (top row, (a) and (b)) and GFLS (bottom row, (c) and (d)).
    Fig. 1  Phase portraits of GFLVS (top row, (a) and (b)) and GFLS (bottom row, (c) and (d)).

    Similarly, we define the generalized fractional Lorenz system (GFLS) as

    $ \begin{align} \begin{cases} D^{\alpha_1}_{0+;[\sigma, w]}x = r(y-x)\\ D^{\alpha_2}_{0+;[\sigma, w]}y = x(\rho-z)-y\\ D^{\alpha_3}_{0+;[\sigma, w]}z = xy-\beta{z} \end{cases} \end{align} $

    (13)

    where $r$ is the Prandtl number, $\rho$ is the Rayleigh number and $\beta$ is the size of the region approximated by the system. The fractional order $0 < \alpha_1, \alpha_2, \alpha_3 < 1$ may take different values.

    By taking the parameters $r=10$ , $\rho=28$ , $\beta= {8}/{3}$ , and the initial condition $[x_0, y_0, z_0]=[0.5, 0.5, 0.5]$ , when $\alpha_1$ $=$ $\alpha_2=\alpha_3=0.99$ , (13) represents the generalized fractional Lorenz chaotic system and the phase portraits of the system (13) are described through Figs. 1(c) and 1(d). In Fig. 1(c), the chaotic attractor of fractional Lorenz system is presented. When we take scale function as a power function, and weight function as exponential function, the GFLS remains chaotic. However, the shape of the attractor changes, which is shown in Fig. 1(d).

    Now we analyze the influence of the scale and weight functions on the responses of generalized fractional differential equation. For simplicity, we consider

    $ \begin{align} D^{\alpha}_{0+;[\sigma, w]}u(t) = Au(t) + f(t) \end{align} $

    (14)

    where $A\neq{0}$ is a constant.

    Equation (14) is equivalent to

    $ \begin{align} \frac{[w(t)]^{-1}}{\Gamma(1-\alpha)}\int^{t}_{0} \frac{[w(s)u(s)]'}{[\sigma(t)-\sigma(s)]^{\alpha}}ds = Au(t) + f(t). \end{align} $

    (15)

    Let $v(t) = w(t)u(t)$ , we have

    $ \begin{align} \frac{1}{\Gamma(1-\alpha)}\int^{t}_{0} \frac{v(s)'}{[\sigma(t)-\sigma(s)]^{\alpha}}ds = Av(t) + w(t)f(t). \end{align} $

    (16)

    According to [13], we deduce the solution of (16) as:

    $ \begin{align} v(t) =&\ E_{\alpha}\left(A[\sigma(t)-\sigma(0)]^{\alpha}\right)v_0 \nonumber\\ & +\int^t_0(\sigma(t)-\sigma(s))^{\alpha-1} \nonumber\\ &\times E_{\alpha, \alpha}[A(\sigma(t)-\sigma(s))^{\alpha}]w(s)f(s)ds \end{align} $

    (17)

    which implies that

    $ \begin{align} u(t)=&\ \frac{w(0)}{w(t)}E_{\alpha}\left(A[\sigma(t)-\sigma(0)]^{\alpha}\right)u_0 \nonumber\\ & +\frac{1}{w(t)}\int^t_0(\sigma(t)-\sigma(s))^{\alpha-1} \nonumber\\ & \times E_{\alpha, \alpha}[A(\sigma(t)-\sigma(s))^{\alpha}]w(s)f(s)ds \end{align} $

    (18)

    where $u_0$ is the initial condition, and $E$ is the Mittag-Leffler function.

    In (18), we observe that how the weight and scale functions influence the behavior of (14). First of all, the weight function cannot be zero in the domain, otherwise solution $u(t)$ will go to infinity. Second, the scale function cannot be periodic, and if it is, the generalized fractional derivative will be infinity at $t=s$ . For an intuitive comprehension, we present some numerical simulations in the following.

    The fractional chaotic systems are sufficiently generalized by using the generalized fractional derivative, since many existing fractional derivatives, as well as integer order derivatives, are special cases of the generalized fractional derivative. In our numerical experiments, we find many interesting dynamical behaviors of generalized fractional chaotic systems which are never found in common fractional or integer order chaotic systems. Here we present some particular simulation results. However, our discussion depends on Figs. 2 and 3, and others figures are not shown here.

    图 2  Influence of scale function $\sigma(t)$ on GFLVS (top row, (a) and (b)) and GFLS (bottom row, (c) and (d)).
    Fig. 2  Influence of scale function $\sigma(t)$ on GFLVS (top row, (a) and (b)) and GFLS (bottom row, (c) and (d)).
    图 3  Influence of weight function $w(t)$ on GFLVS (top row, (a) and (b)) and GFLS (bottom row, (c) and (d)).
    Fig. 3  Influence of weight function $w(t)$ on GFLVS (top row, (a) and (b)) and GFLS (bottom row, (c) and (d)).

    First, we simulate the influence of scale function on dynamics of chaotic systems. In GFLVS, we take fractional order $\alpha_1=\alpha_2=\alpha_3=0.95$ , weight function $w(t)=\exp(1.2t)$ , and other parameters are the same as before. In GFLS, we select fractional order $\alpha_1=\alpha_2=\alpha_3=0.99$ , weight function $w(t)=\exp(0.1t)$ , and other parameters are the same as before. The dynamic behaviors of GFLVS and GFLS with scale function $\sigma(t)=t$ and $t^{1.14}$ are individually presented in Fig. 2.

    Second, we simulate the influence of weight function on dynamics of chaotic systems. In GFLVS, we take fractional order $\alpha_1=\alpha_2=\alpha_3=0.95$ , scale function $\sigma(t)=t$ , and other parameters are the same as before. In GFLS, we select fractional order $\alpha_1=\alpha_2=\alpha_3=0.99$ , scale function $\sigma(t)=t$ , and other parameters are the same as before. The dynamic behaviors of GFLVS with weight function $w(t)$ $=$ $\exp(0.8t)$ , $\exp(1.3t)$ , and GFLS with weight function $w(t)=\exp(2+0.5t)$ and $\exp(2+0.2t)$ are presented in Fig. 3.

    Finally, to end this section, we make some remarks based on the numerical experiments above. Some other figures are not listed here for shortening the length of paper.

    1) The GFLVS is chaotic with scale function $\sigma(t)=t$ , weight function $w(t)$ is a nonzero constant, and fractional order $\alpha_i=0.95$ , $i=1, 2, 3$ [27]. However, From Fig. 1(a), Fig. 2(a) and Fig. 3(a), we may see that as the weight function varies, the chaotic attractor vanishes and then a limit cycle emerges or the system converges to a stable equilibrium point. Furthermore, from Fig. 2(a) and Fig. 2(b), we observe that as the scale function varies, the limit cycle tends to be a stable equilibrium point. From Fig. 3(a) and Fig. 3(b), it is shown that as the weight function varies, the limit cycle can be generated from a stable equilibrium point.

    2) The GFLS is chaotic with scale function $\sigma(t)=t$ , weight function $w(t)$ is a nonzero constant, and fractional order $\sum^{3}_{i=1}\alpha_i>2.91$ [28]. In simulation, on one hand, Figs. 1(c) and 1(d), indicate that with suitable scale and weight functions, the GFLS also has a chaotic attractor. On the other hand, Fig. 1(c), Fig. 1(d), Fig. 2(c), Fig. 2(d), and Fig. 3(d) imply that the scale and weight functions can influence the shape and position of chaotic attractor. From Figs. 3(c) and 3(d), we observe that with some suitable weight function, the chaotic attractor tends to be an asymptotically stable equilibrium point.

    3) Our previous work [23]-[26] verified that in generalized fractional integral and generalized fractional derivative, the basic property of scale function $\sigma(t)$ is that it changes the time axis, which means that if the time domain is specified as $[0, T]$ , then the response of the dynamical system is obtained over $[\sigma(0), \sigma(T)]$ , provided the scale function is monotone increasing. Since the chaotic dynamical systems are sensitive to the initial conditions, when we take different scale functions in generalized fractional chaotic system, many different dynamical behaviors will be drawn.

    4) A similar observation to weight function can be found in [23]-[26], which shows that in generalized fractional integral and generalized fractional derivative, the basic property of weight function $w(t)$ is that it puts different weights for function in different positions of domain. The classical fractional operators have memory property which makes them excellent tools to model the diffusion process with heredity. Generally, in left Caputo type generalized fractional derivative, the monotonic increasing weight function is coincident with the inner memory property of fractional operator, while the monotonic decreasing weight function can destroy this inner property. One can also follow our numerical method and try other scale and weight functions in numerical experiments.

    5) In Figs. 2 and 3, one can observe that both changing the scale and weight functions make the systems change between different dynamical behavior (e.g., limit cycle and stable equilibrium point). These phenomena can be regarded as general cases for generalized fractional chaotic systems. We shall guess that either scale function or weight function would influence the dynamics of generalized fractional chaotic systems. In Fig. 2, the weight function is fixed so that the influence of scale function on GFLVS and GFLS is presented. Similarly, in Fig. 3, the scale function is fixed so that the influence of weight function on GFLVS and GFLS is shown. From (18), we clearly see that the scale function plays an important role in scaling the long time behavior of dynamics since it is located in the generalized exponential function, and the weight function provides a different average since it lies inside the integral, and it is a variable coefficient simultaneously. Apparently, the behavior of function $u$ depends on the changing of scale and weight functions.

    In this paper, we presented a class of new generalized fractional chaotic system, using the new generalized fractional derivative proposed recently. Many dynamical systems with integer or fractional order derivatives can be extended by replacing the derivative with the generalized fractional derivative. Therefore, the new generalized fractional dynamical systems considered in this paper can exhibit more complex dynamic behaviors. In simulations, we show that the dynamical behaviors of such systems not only depend on fractional order, but also depend on the scale and weight functions.

    Acknowledgement: The author is grateful to Professor O. P. Agrawal (SIUC, USA) for introducing him theory of generalized fractional calculus, suggesting the basic idea of this paper, as well as his kind help and continuous encouragement in the recent years.
  • 图  1  AlphaGo的分析1

    Fig.  1  Analysis 1 of AlphaGo

    图  2  AlphaGo的分析2

    Fig.  2  Analysis 2 of AlphaGo

    表  1  阿法狗在快速走子中使用的盘面特征

    Table  1  Input features for rollout and tree policy

    Feature # of patterns Description
    Response 1 Whether move matches one or more response pattern features
    Save atari 1 Move saves stone(s) from capture
    Neighbour 8 Move is 8-connected to previous move
    Nakade 8 192 Move matches a nakade pattern at captured stone
    Response pattern 32 207 Move matches 12-point diamond pattern near previous move
    Non-response pattern 69 338 Move matches 3 £ 3 pattern around move
    Self-atari 1 Move allows stones to be captured
    Last move distance 34 Manhattan distance to previous two moves
    Non-response pattern 32 207 Move matches 12-point diamond pattern centred around move
    (Features used by the rollout pollcy (the frst set) and tree policy (the frst and second sets). Patterns are based on stone colour (black/white/empty) and liberties (1, 2,≥3) at each intersection of the pattern.)
    下载: 导出CSV

    表  2  不同版本阿法狗的等级分比较(等级分由一场内部锦标赛决出)

    Table  2  Results of a tournament between di®erent variants of AlphaGo

    Short name Policy network Value network Rollouts Mixing constant Policy GPUs Value GPUs Elo rating
    αrvp pσ vθ pπ λ= 0.5 2 6 2 890
    αvp pσ vθ - λ= 0 2 6 2 177
    αrp pσ - pπ λ= 1 8 0 2 416
    αrv [pτ] vθ pπ λ= 0.5 0 8 2 077
    αv [pτ] vθ - λ= 0 0 8 1 655
    αr [pτ] - pπ λ= 1 0 0 1 457
    αp pσ - - - 0 0 1 517
    Evaluating positions using rollouts only (αrp; αr), value nets only (αvp; αv), or mixing both (αrvp; αrv); either using the policy network ρσ(αrvp; αvp; αrp) or no policy network (αrvp; αvp; αrp), that is, instead using the placeholder probabilities from the tree policy pτ throughout. Each program used 5 s per move on a single machine with 48 CPUs and 8 GPUs. Elo ratings were computed by BayesElo.
    下载: 导出CSV
  • [1] Silver D, Huang A, Maddison C J, Guez A, Sifre L, van den Driessche G, Schrittwieser J, Antonoglou I, Panneershelvam V, Lanctot M, Dieleman S, Grewe D, Nham J, Kalchbrenner N, Sutskever I, Lillicrap T, Leach M, Kavukcuoglu K, Graepel T, Hassabis D. Mastering the game of go with deep neural networks and tree search. Nature, 2016, 529(7587): 484-489
    [2] Tian Y D, Zhu Y. Better computer go player with neural network and long-term prediction. In: International Conference on Learning Representation (ICLR). San Juan, Puerto Rico, 2016.
  • 期刊类型引用(3)

    1. 曾超,王文军,陈朝阳,张超飞,成波. 模拟驾驶实验中同步装置设计及实验教学应用. 实验技术与管理. 2020(03): 99-102 . 百度学术
    2. 吴思凡,杜煜,徐世杰,杨硕,杜晨. 基于长短期记忆-异步优势动作评判的智能车汇入模型. 汽车技术. 2019(10): 42-47 . 百度学术
    3. 成英,高利,陈雪梅,赵亚男. 有人与无人驾驶车辆交叉口驾驶博弈模型. 北京理工大学学报. 2019(09): 938-943 . 百度学术

    其他类型引用(17)

  • 加载中
图(2) / 表(2)
计量
  • 文章访问数:  4350
  • HTML全文浏览量:  897
  • PDF下载量:  4902
  • 被引次数: 20
出版历程
  • 收稿日期:  2016-04-14
  • 录用日期:  2016-05-10
  • 刊出日期:  2016-05-01

目录

/

返回文章
返回