Abstract
In this paper, we propose a software tool, called AMYTISS, implemented in C++/OpenCL, for designing correctbyconstruction controllers for largescale discretetime stochastic systems. This tool is employed to (i) build finite Markov decision processes (MDPs) as finite abstractions of given original systems, and (ii) synthesize controllers for the constructed finite MDPs satisfying boundedtime highlevel properties including safety, reachability and reachavoid specifications. In AMYTISS, scalable parallel algorithms are designed such that they support the parallel execution within CPUs, GPUs and hardware accelerators (HWAs). Unlike all existing tools for stochastic systems, AMYTISS can utilize highperformance computing (HPC) platforms and cloudcomputing services to mitigate the effects of the stateexplosion problem, which is always present in analyzing largescale stochastic systems. We benchmark AMYTISS against the most recent tools in the literature using several physical case studies including robot examples, room temperature and road traffic networks. We also apply our algorithms to a 3dimensional autonomous vehicle and 7dimensional nonlinear model of a BMW 320i car by synthesizing an autonomous parking controller.
Keywords
A. Lavaei and M. Khaled—Authors have contributed equally.
This work was supported in part by the H2020 ERC Starting Grant AutoCPS (grant agreement No. 804639).
Download conference paper PDF
1 Introduction
1.1 Motivations
Largescale stochastic systems are an important modeling framework to describe many reallife safetycritical systems such as power grids, traffic networks, selfdriving cars, and many other applications. For this type of complex systems, automating the controller synthesis procedure to achieve highlevel specifications, e.g., those expressed as linear temporal logic (LTL) formulae [24], is inherently very challenging mainly due to their computational complexity arising from uncountable sets of states and actions. To mitigate the encountered difficulty, finite abstractions, i.e., systems with finite state sets, are usually employed as replacements of original continuousspace systems in the controller synthesis procedure. More precisely, one can first abstract a given continuousspace system by a simpler one, e.g., a finite Markov decision process (MDP), and then perform analysis and synthesis over the abstract model (using algorithmic techniques from computer science [3]). Finally, the results are carried back to the original system, while providing a guaranteed error bound [5, 13,14,15,16,17,18,19,20,21, 23].
Unfortunately, construction of finite MDPs for largescale complex systems suffers severely from the socalled curse of dimensionality: the computational complexity grows exponentially as the number of state variables increases. To alleviate this issue, one promising solution is to employ highperformance computing (HPC) platforms together with cloudcomputing services to mitigate the stateexplosion problem. In particular, HPC platforms have a large number of processing elements (PEs) and this significantly affects the time complexity when serial algorithms are parallelized [7].
1.2 Contributions
The main contributions and merits of this work are:

(1)
We propose a novel dataparallel algorithm for constructing finite MDPs from discretetime stochastic systems and storing them in efficient distributed data containers. The proposed algorithm handles largescale systems.

(2)
We propose a parallel algorithm for synthesizing discrete controllers using the constructed MDPs to satisfy safety, reachability, or reachavoid specifications. More specifically, we introduce a parallel algorithm for the iterative computation of Bellman equation in standard dynamic programming [26, 27].

(3)
Unlike the existing tools in the literature, AMYTISS accepts bounded disturbances and natively supports both additive and multiplicative noises with different practical distributions including normal, uniform, exponential, and beta.
We apply the proposed implementations to realworld applications including robot examples, room temperature and road traffic networks, and autonomous vehicles. This extends the applicability of formal methods to some safetycritical realworld applications with high dimensions. The results show remarkable reductions in the memory usage and computation time outperforming all existing tools in the literature.
We provide AMYTISS as an opensource tool. After compilation, AMYTISS is loaded via pFaces [10] and launched for parallel execution within available parallel computing resources. The source of AMYTISS and detailed instructions on its building and running can be found in: https://github.com/mkhaled87/pFacesAMYTISS
Due to lack of space, we provide details of traditional serial and proposed parallel algorithms, case studies, etc. in an arXiv version of the paper [12].
1.3 Related Literature
There exist several software tools on verification and synthesis of stochastic systems with different classes of models. SReachTools [30] performs stochastic reachability analysis for linear, potentially timevarying, discretetime stochastic systems. ProbReach [25] is a tool for verifying the probabilistic reachability for stochastic hybrid systems. SReach [31] solves probabilistic bounded reachability problems for two classes of models: (i) nonlinear hybrid automata with parametric uncertainty, and (ii) probabilistic hybrid automata with additional randomness for both transition probabilities and variable resets. Modest Toolset [6] performs modeling and analysis for hybrid, realtime, distributed and stochastic systems. Two competitions on tools for formal verification and policy synthesis of stochastic models are organized with reports in [1, 2].
FAUST\(^{\mathsf 2}\) [29] generates formal abstractions for continuousspace discretetime stochastic processes, and performs verification and synthesis for safety and reachability specifications. However, FAUST\(^{\mathsf 2}\) is originally implemented in MATLAB and suffers from the curse of dimensionality due to its lack of scalability for largescale models. StocHy [4] provides the quantitative analysis of discretetime stochastic hybrid systems such that it constructs finite abstractions, and performs verification and synthesis for safety and reachability specifications.
AMYTISS differs from FAUST\(^{\mathsf 2}\) and StocHy in two main directions. First, AMYTISS implements novel parallel algorithms and data structures targeting HPC platforms to reduce the undesirable effects of the stateexplosion problem. Accordingly, it is able to perform parallel execution in different heterogeneous computing platforms including CPUs, GPUs and HWAs. Whereas, FAUST\(^{\mathsf 2}\) and StocHy can only run serially on one CPU, and consequently, it is limited to small systems. Additionally, AMYTISS can handle the abstraction construction and controller synthesis for two and a half player games (e.g., stochastic systems with bounded disturbances), whereas FAUST\(^{\mathsf 2}\) and StocHy only handle one and a half player games (e.g., disturbancefree systems).
Unlike all existing tools, AMYTISS offers highly scalable, distributed execution of parallel algorithms utilizing all available processing elements (PEs) in any heterogeneous computing platform. To the best of our knowledge, AMYTISS is the only tool of its kind for continuousspace stochastic systems that is able to utilize all types of compute units (CUs), simultaneously.
We compare AMYTISS with FAUST\(^{\mathsf 2}\) and StocHy in Table 1 in detail in terms of different technical aspects. Although there have been some efforts in FAUST\(^{\mathsf 2}\) and StocHy for parallel implementations, these are not compatible with HPC platforms. Specifically, FAUST\(^{\mathsf 2}\) employs some parallelization techniques using parallel forloops and sparse matrices inside Matlab, and StocHy uses Armadillo, a multithreaded library for scientific computing. However, these tools are not designed for the parallel computation on HPC platforms. Consequently, they can only utilize CPUs and cannot run on GPUs or HWAs. In comparison, AMYTISS is developed in OpenCL, a language specially designed for dataparallel tasks, and supports heterogeneous computing platforms combining CPUs, GPUs and HWAs.
Note that FAUST\(^{\mathsf 2}\) and StocHy do not natively support reachavoid specifications in the sense that users can explicitly provide some avoid sets. Implementing this type of properties requires some modifications inside those tools. In addition, we do not make a comparison here with SReachTools since it is mainly for stochastic reachability analysis of linear, potentially timevarying, discretetime stochastic systems, while AMYTISS is not limited to reachability analysis and can handle nonlinear systems as well.
Note that we also provide a script in the tool repository^{Footnote 1} that converts the MDPs constructed by AMYTISS into PRISMinputfiles [11]. In particular, AMYTISS can natively construct finite MDPs from continuousspace stochastic control systems. PRISM can then be employed to perform the controller synthesis for those classes of complex specifications that AMYTISS does not support.
2 DiscreteTime Stochastic Control Systems
We formally introduce discretetime stochastic control systems (dtSCS) below.
Definition 1
A discretetime stochastic control system (dtSCS) is a tuple
where,

\(X\!\subseteq \!\mathbb R^n\) is a Borel space as the state set and \((\!X, \!\mathcal B (X)\!)\) is its measurable space;

\(U\!\subseteq \! \mathbb R^m\) is a Borel space as the input set;

\(W\!\subseteq \! \mathbb R^p\) is a Borel space as the disturbance set;

\(\varsigma \) is a sequence of independent and identically distributed (i.i.d.) random variables from a sample space \(\varOmega \) to a measurable set \(\mathcal V_\varsigma \)
$$\begin{aligned} \varsigma :=\{\varsigma (k):\varOmega \rightarrow \mathcal V_{\varsigma },\,\,k\in {\mathbb {N}}\}; \end{aligned}$$ 
\(f:X\times U\times W\rightarrow X\) is a measurable function characterizing the state evolution of the system.
The state evolution of \(\varSigma \), for a given initial state \(x(0)\in X\), an input sequence \(\nu (\cdot ):\mathbb N\rightarrow U\), and a disturbance sequence \(w(\cdot ):\mathbb N\rightarrow W\), is characterized by the difference equations
where \(\varUpsilon (k) := \varsigma (k)\) with \(\mathcal V_\varsigma = \mathbb R^n\) for the case of the additive noise, and \(\varUpsilon (k) := \varsigma (k)x(k)\) with \(\mathcal V_\varsigma \) equals to the set of diagonal matrices of the dimension n for the case of the multiplicative noise [22]. We keep the notation \(\varSigma \) to indicate both cases and use respectively \(\varSigma _\mathfrak a\) and \(\varSigma _\mathfrak m\) when discussing these cases individually.
We should mention that our parallel algorithms are independent of the noise distribution. For an easier presentation of the contribution, we present our algorithms and case studies based on normal distributions but our tool natively supports other practical distributions including uniform, exponential, and beta. In addition, we provide a subroutine in our software tool so that the user can still employ the parallel algorithms by providing the density function of the desired class of distributions.
Remark 1
Our synthesis is based on a \(\max \)\(\min \) optimization problem for two and a half player games by considering the disturbance and input of the system as players [9]. Particularly, we consider the disturbance affecting the system as an adversary and maximize the probability of satisfaction under the worstcase strategy of a rational adversary. Hence, we minimize the probability of satisfaction with respect to disturbances, and maximize it over control inputs.
One may be interested in analyzing dtSCSs without disturbances (cf. case studies). In this case, the tuple (1) reduces to \(\varSigma =(X,U,\varsigma ,f)\), where \(f:X\times U\rightarrow X\), and the Eq. (2) can be rewritten as
Note that input models in this tool paper are given inside configuration text files. Systems are described by stochastic difference equations as (2)–(3), and the user should provide the righthandside of equations^{Footnote 2}. In the next section, we formally define MDPs and discuss how to build finite MDPs from given dtSCSs.
3 Finite Markov Decision Processes (MDPs)
A dtSCS \(\varSigma \) in (1) is equivalently represented by the following MDP [8, Proposition 7.6]:
where the map \(T_{\mathsf x}:\mathcal B(X)\times X\times U\times W\rightarrow [0,1]\), is a conditional stochastic kernel that assigns to any \(x \in X\), \(\nu \in U\), and \(w\in W\), a probability measure \(T_{\mathsf x}(\cdot  x,\nu , w)\). The alternative representation as the MDP is utilized in [28] to approximate a dtSCS \(\varSigma \) with a finite MDP \(\widehat{\varSigma }\) using an abstraction algorithm. This algorithm first constructs a finite partition of the state set \(X = \cup _i \mathsf X_i\), the input set \(U = \cup _i \mathsf U_i\), and the disturbance set \(W = \cup _i \mathsf W_i\). Then representative points \(\bar{x}_i\in \mathsf X_i\), \(\bar{\nu }_i\in \mathsf U_i\), and \(\bar{w}_i\in \mathsf W_i\) are selected as abstract states, inputs, and disturbances. The transition probability matrix for the finite MDP \(\widehat{\varSigma }\) is also computed as
where the map \(\varXi :X\rightarrow 2^X\) assigns to any \(x\in X\), the corresponding partition element it belongs to, i.e., \(\varXi (x) = \mathsf X_i\) if \(x\in \mathsf X_i\). Since \(\hat{X}\), \(\hat{U}\) and \(\hat{W}\) are finite sets, \(\hat{T}_{\mathsf x}\) is a static map. It can be represented with a matrix and we refer to it, from now on, as the transition probability matrix.
For a given logic specification \(\varphi \) and accuracy level \(\epsilon \), the discretization parameter \(\delta \) can be selected a priori such that
where \(\epsilon \) depends on the horizon of formula \(\varphi \), the Lipschitz constant of the stochastic kernel, and the state discretization parameter \(\delta \) (cf. [28, Theorem 9]). We refer the interested reader to the arXiv version [12] for more details.
In the next sections, we propose novel parallel algorithms for the construction of finite MDPs and the synthesis of their controllers.
4 Parallel Construction of Finite MDPs
In this section, we propose an approach to efficiently compute the transition probability matrix \(\hat{T}_{\mathsf x}\) of the finite MDP \(\widehat{\varSigma }\), which is essential for any controller synthesis procedure, as we discuss later in Sect. 5.
4.1 DataParallel Threads for Computing \(\hat{T}_{\mathsf X}\)
The serial algorithm for computing \(\hat{T}_{\mathsf x}\) is presented in Algorithm 1 in the arXiv version [12]. Computations of mean \(\mu = f(\bar{x}_i,\bar{\nu }_j,\bar{w}_k, 0)\), \(\text {PDF}(x\,\,\mu , \mathsf {\varSigma })\), where PDF stands for probability density functions and \(\mathsf {\varSigma }\) is a noise covariance matrix, and of \(\hat{T}_{\mathsf x}\) all do not share data from one innerloop to another. Hence, this is an embarrassingly dataparallel section of the algorithm. pFaces [10] can be utilized to launch necessary number of parallel threads on the employed hardware configuration (HWC) to improve the computation time of the algorithm. Each thread will eventually compute and store, independently, its corresponding values within \(\hat{T}_{\mathsf x}\).
4.2 Less Memory for Post States in \(\hat{T}_{\mathsf X}\)
\(\hat{T}_{\mathsf x}\) is a matrix with the dimension of \((n_x\times n_\nu \times n_w,n_x)\). The number of columns is \(n_x\) as we need to compute and store the probability for each reachable partition element \(\varXi (x'_l)\), corresponding to the representing post state \(x'_l\). Here, we consider the Gaussian PDFs for the sake of a simpler presentation. For simplicity, we now focus on the computation of tuple \((\bar{x}_i, \bar{\nu }_j, \bar{w}_k)\). In many cases, when the PDF is decaying fast, only partition elements near \(\mu \) have high probabilities of being reached, starting from \(\bar{x}_i\) and applying an input \(\bar{\nu }_j\).
We set a cutting probability threshold \(\gamma \in [0,1]\) to control how many partition elements around \(\mu \) should be stored. For a given mean value \(\mu \), a covariance matrix \(\mathsf {\varSigma }\) and a cutting probability threshold \(\gamma \), \(x \in X\) is called a PDF cutting point if \(\gamma = \text {PDF}(x \vert \mu , \mathsf {\varSigma })\). Since Gaussian PDFs are symmetric, by repeating this cutting process dimensionwise, we end up with a set of points forming a hyperrectangle in X, which we call it the cutting region and denote it by \(\hat{X}^{\mathsf {\varSigma }}_{\gamma }\). This is visualized in Fig. 1 in the arXiv version [12] for a 2dimensional system. Any partition element \(\varXi (x'_l)\) with \(x'_l\) outside the cutting region is considered to have zero probability of being reached. Such approximation allows controlling the sparsity of the columns of \(\hat{T}_{\mathsf x}\). The closer the value of \(\gamma \) to zero, the more accurate \(\hat{T}_{\mathsf x}\) in representing transitions of \(\widehat{\varSigma }\). On the other hand, the closer the value of \(\gamma \) to one, less post state values need to be stored as columns in \(\hat{T}_{\mathsf x}\). The number of probabilities to be stored for each \((\bar{x}_i, \bar{\nu }_j, \bar{w}_k)\) is then \(\vert \hat{X}^{\mathsf {\varSigma }}_{\gamma } \vert \).
Note that since \(\mathsf {\varSigma }\) is fixed prior to running the algorithm, number of columns needed for a fixed \(\gamma \) can be identified before launching the computation. We can then accurately allocate a uniform fixed number of memory locations for any tuple \((\bar{x}_i, \bar{\nu }_j, \bar{w}_k)\) in \(\hat{T}_{\mathsf x}\). Hence, there is no need for a dynamic sparse matrix data structure and \(\hat{T}_{\mathsf x}\) is now a matrix with a dimension of \((n_x\times n_\nu \times n_w,\vert \hat{X}^{\mathsf {\varSigma }}_{\gamma } \vert )\).
4.3 A Parallel Algorithm for Constructing Finite MDP \(\widehat{\varSigma }\)
We present a novel parallel algorithm (Algorithm 2 in the arXiv version [12]) to efficiently construct and store \(\hat{T}_{\mathsf x}\) as a successor. We employ the discussed enhancements in Subsect. 4.1 and 4.2 within the proposed algorithm. We do not parallelize the forloop in Algorithm 2, Step 2, to avoid excessive parallelism (i.e., we parallelize loops only over X and U, but not over W). Note that, practically, for largescale systems, \(\vert \hat{X} \times \hat{U} \vert \) can reach up to billions. We are interested in the number of parallel threads that can be scheduled reasonably by available HW computing units.
5 Parallel Synthesis of Controllers
In this section, we employ dynamic programming to synthesize controllers for constructed finite MDPs \(\widehat{\varSigma }\) satisfying safety, reachability, and reachavoid properties [26, 27]. The classical serial algorithm and its proposed parallelized version are respectively presented as Algorithms 3 and 4 in the arXiv version [12]. We should highlight that the parallelism here mainly comes from the parallelization of matrix multiplication and the loop over timesteps cannot be parallelized due to the data dependency. More details can be found in the arXiv version.
5.1 OntheFly Construction of \(\hat{T}_{\mathsf X}\)
In AMYTISS, we also use another technique that further reduces the required memory for computing \(\hat{T}_{\mathsf x}\). We refer to this approach as onthefly abstractions (OFA). In OFA version of Algorithm 4 [12], we skip computing and storing the MDP \(\hat{T}_{\mathsf x}\) and the matrix \(\hat{T}_{0\mathsf x}\) (i.e., Steps 1 and 5). We instead compute the required entries of \(\hat{T}_{\mathsf x}\) and \(\hat{T}_{0\mathsf x}\) onthefly as they are needed (i.e., Steps 13 and 15). This significantly reduces the required memory for \(\hat{T}_{\mathsf x}\) and \(\hat{T}_{0\mathsf x}\) but at the cost of repeated computation of their entries in each time step from 1 to \(T_d\). This gives the user an additional control over the tradeoff between the computation time and memory.
5.2 Supporting Multiplicative Noises and Practical Distributions
AMYTISS natively supports multiplicative noises and practical distributions such as uniform, exponential, and beta distributions. The technique introduced in Subsect. 4.2 for reducing the memory usage is also tuned for other distributions based on the support of their PDFs. Since AMYTISS is designed for extensibility, it allows also for customized distributions. Users need to specify their desired PDFs and hyperrectangles enclosing their supports so that AMYTISS can include them in the parallel computation of \(\hat{T}_{\mathsf x}\). Further details on specifying customized distributions are provided in the README file.
AMYTISS also supports multiplicative noises as introduced in (2). Currently, the memory reduction technique of Subsect. 4.2 is disabled for systems with multiplicative noises. This means users should expect larger memory requirements for systems with multiplicative noises. However, users can still benefit from the proposed OFA version to compensate for the increase in memory requirement. We plan to include this feature for multiplicative noises in a future update of AMYTISS. Note that for a better demonstration, previous sections were presented by the additive noise and Gaussian normal PDF to introduce the concepts.
6 Benchmarking and Case Studies
AMYTISS is selfcontained and requires only a modern C++ compiler. It supports all major operating systems: Windows, Linux and Mac OS. Once compiled, utilizing AMYTISS is a matter of providing text configuration files and launching the tool. AMYTISS implements scalable parallel algorithms that run on top of pFaces [10]. Hence, users can utilize computing power in HPC platforms and cloud computing to scale the computation and control the computational complexities of their problems. Table 2 lists the HW configuration we use to benchmark AMYTISS. The devices range from local devices in desktop computers to advanced compute devices in Amazon AWS cloud computing services.
Table 3 shows the benchmarking results running AMYTISS with these HWCs for several case studies and makes comparisons between AMYTISS, FAUST\(^{\mathsf 2}\), and StocHy. We employ a machine with Windows operating system (Intel i7@3.6 GHz CPU and 16 GB of RAM) for FAUST\(^{\mathsf 2}\), and StocHy. It should be mentioned that FAUST\(^{\mathsf 2}\) predefines a minimum number of representative points based on the desired abstraction error, and accordingly the computation time and memory usage reported in Table 3 are based on the minimum number of representative points. In addition, to have a fair comparison, we run all the case studies with additive noises since neither FAUST\(^{\mathsf 2}\) nor StocHy supports multiplicative noises.
To show the applicability of our results to largescale systems, we apply our techniques to several physical case studies. We synthesize controllers for 3 and 5dimensional room temperature networks to keep temperatures in a comfort zone. Furthermore, we synthesize controllers for road traffic networks with 3 and 5 dimensions to keep the density of the traffic below some desired level. In addition, we apply our algorithms to a 2dimensional nonlinear robot and synthesize controllers satisfying safety and reachavoid specifications. Finally, we consider 3 and 7dimensional nonlinear models of an autonomous vehicle and synthesize reachavoid controllers to automatically park the vehicles. For details of case studies, see the arXiv version [12].
Table 3 presents a comparison between AMYTISS, FAUST\(^{\mathsf 2}\) and StocHy w.r.t the computation time and required memory. For each HWC, we show the time in seconds to solve the problem. Clearly, employing HWCs with more PEs reduces the time to solve the problem. This is a strong indication for the scalability of the proposed algorithms. Since AMYTISS is the only tool for stochastic systems that can utilize the reported HWCs, we do not compare it with other similar tools.
In Table 3, first 13 rows, we also include the benchmark provided in StocHy [4, Case study 3]. Table 4 in the arXiv version [12] shows an additional comparison between StocHy and AMYTISS on a machine with the same configuration as the one employed in [4] (a laptop having an Intel Core i\(78550\)U CPU at 1.80GHz with 8 GB of RAM). StocHy suffers significantly from the stateexplosion problem as seen from its exponentially growing computation time. AMYTISS, on the other hand, outperforms StocHy and can handle bigger systems using the same hardware.
As seen in Table 3, AMYTISS outperforms FAUST\(^{\mathsf 2}\) and StocHy in all the case studies (maximum speedups up to 692000 times). Moreover, AMYTISS is the only tool that can utilize the available HW resources. The OFA feature in AMYTISS reduces dramatically the required memory, while still solves the problems in a reasonable time. FAUST\(^{\mathsf 2}\) and StocHy fail to solve many of the problems since they lack the native support for nonlinear systems, they require large amounts of memory, or they do not finish computing within 24 hours.
Note that considering only dimensions of systems can be sometimes misleading. In fact, number of transitions in MDPs (\(\vert \hat{X} \times \hat{U} \vert \)) can give a better judgment on the size of systems since it directly affects the memory/time needed for solving the problem. For instance in Table 3, the number of transitions for the 14dimensional case study is 16384, while for the 5dimensional room temperature example is 279936 transitions (i.e., almost 17 times bigger). This means AMYTISS can clearly handle much larger systems than existing tools.
Notes
 1.
 2.
An example of such a configuration file is provided at: https://github.com/mkhaled 87/pFacesAMYTISS/blob/master/examples/extoysafety/toy2d.cfg.
References
Abate, A., et al.: ARCHCOMP19 category report: stochastic modelling. EPiC Ser. Comput. 61, 62–102 (2019)
Abate, A., et al.: ARCHCOMP18 category report: Stochastic modelling. In: ARCH@ ADHS, pp. 71–103 (2018)
Baier, C., Katoen, J.P.: Principles of Model Checking. MIT Press, Cambridge (2008)
Cauchi, N., Abate, A.: \(\sf StocHy\): automated verification and synthesis of stochastic processes. In: Vojnar, T., Zhang, L. (eds.) TACAS 2019. LNCS, vol. 11428, pp. 247–264. Springer, Cham (2019). https://doi.org/10.1007/9783030174651_14
Haesaert, S., Soudjani, S.: Robust dynamic programming for temporal logic control of stochastic systems. CoRR abs/1811.11445 (2018). http://arxiv.org/abs/1811.11445
Hartmanns, A., Hermanns, H.: The modest toolset: an integrated environment for quantitative modelling and verification. In: Ábrahám, E., Havelund, K. (eds.) TACAS 2014. LNCS, vol. 8413, pp. 593–598. Springer, Heidelberg (2014). https://doi.org/10.1007/9783642548628_51
Jaja, J.: An Introduction to Parallel Algorithms. AddisonWesley, Boston (1992)
Kallenberg, O.: Foundations of Modern Probability. Springer, New York (1997). https://doi.org/10.1007/b98838
Kamgarpour, M., Ding, J., Summers, S., Abate, A., Lygeros, J., Tomlin, C.: Discrete time stochastic hybrid dynamical games: Verification & controller synthesis. In: Proceedings of the 50th IEEE Conference on Decision and Control and European Control Conference, pp. 6122–6127 (2011)
Khaled, M., Zamani, M.: pFaces: an acceleration ecosystem for symbolic control. In: Proceedings of the 22nd ACM International Conference on Hybrid Systems: Computation and Control, pp. 252–257 (2019)
Kwiatkowska, M., Norman, G., Parker, D.: PRISM: probabilistic symbolic model checker. In: Field, T., Harrison, P.G., Bradley, J., Harder, U. (eds.) TOOLS 2002. LNCS, vol. 2324, pp. 200–204. Springer, Heidelberg (2002). https://doi.org/10.1007/3540460292_13
Lavaei, A., Khaled, M., Soudjani, S., Zamani, M.: AMYTISS: parallelized automated controller synthesis for largescale stochastic system. arXiv:2005.06191, May 2020
Lavaei, A., Soudjani, S., Majumdar, R., Zamani, M.: Compositional abstractions of interconnected discretetime stochastic control systems. In: Proceedings of the 56th IEEE Conference on Decision and Control, pp. 3551–3556 (2017)
Lavaei, A., Soudjani, S., Zamani, M.: Compositional synthesis of finite abstractions for continuousspace stochastic control systems: a smallgain approach. In: Proceedings of the 6th IFAC Conference on Analysis and Design of Hybrid Systems, vol. 51, pp. 265–270 (2018)
Lavaei, A., Soudjani, S., Zamani, M.: From dissipativity theory to compositional construction of finite Markov decision processes. In: Proceedings of the 21st ACM International Conference on Hybrid Systems: Computation and Control, pp. 21–30 (2018)
Lavaei, A., Soudjani, S., Zamani, M.: Compositional abstractionbased synthesis of general MDPs via approximate probabilistic relations. arXiv: 1906.02930 (2019)
Lavaei, A., Soudjani, S., Zamani, M.: Compositional construction of infinite abstractions for networks of stochastic control systems. Automatica 107, 125–137 (2019)
Lavaei, A., Soudjani, S., Zamani, M.: Compositional abstractionbased synthesis for networks of stochastic switched systems. Automatica 114, 108827 (2020)
Lavaei, A., Soudjani, S., Zamani, M.: Compositional abstraction of largescale stochastic systems: a relaxed dissipativity approach. Nonlinear Anal. Hybrid Syst. 36, 100880 (2020)
Lavaei, A., Soudjani, S., Zamani, M.: Compositional (in)finite abstractions for largescale interconnected stochastic systems. IEEE Trans. Autom. Control. (2020). https://doi.org/10.1109/TAC.2020.2975812
Lavaei, A., Zamani, M.: Compositional construction of finite MDPs for largescale stochastic switched systems: a dissipativity approach. In: Proceedings of the 15th IFAC Symposium on Large Scale Complex Systems: Theory and Applications 52(3), 31–36 (2019)
Li, W., Todorov, E., Skelton, R.E.: Estimation and control of systems with multiplicative noise via linear matrix inequalities. In: Proceedings of the American Control Conference, pp. 1811–1816 (2005)
Mallik, K., Schmuck, A., Soudjani, S., Majumdar, R.: Compositional synthesis of finitestate abstractions. IEEE Trans. Autom. Control. 64(6), 2629–2636 (2019)
Pnueli, A.: The temporal logic of programs. In: Proceedings of the 18th Annual Symposium on Foundations of Computer Science, pp. 46–57 (1977)
Shmarov, F., Zuliani, P.: ProbReach: verified probabilistic deltareachability for stochastic hybrid systems. In: Proceedings of the 18th International Conference on Hybrid Systems: Computation and Control, pp. 134–139 (2015)
Soudjani, S.: Formal abstractions for automated verification and synthesis of stochastic systems. Ph.D. thesis, Technische Universiteit Delft, The Netherlands (2014)
Soudjani, S., Abate, A.: Adaptive and sequential gridding procedures for the abstraction and verification of stochastic processes. SIAM J. Appl. Dyn. Syst. 12(2), 921–956 (2013)
Soudjani, S., Abate, A., Majumdar, R.: Dynamic Bayesian networks as formal abstractions of structured stochastic processes. In: Proceedings of the 26th International Conference on Concurrency Theory, pp. 1–14 (2015)
Soudjani, S.E.Z., Gevaerts, C., Abate, A.: FAUST\(^\mathsf{2}\): Formal Abstractions of UncountableSTate STochastic Processes. In: Baier, C., Tinelli, C. (eds.) TACAS 2015. LNCS, vol. 9035, pp. 272–286. Springer, Heidelberg (2015). https://doi.org/10.1007/9783662466810_23
Vinod, A.P., Gleason, J.D., Oishi, M.M.: SReachTools: a MATLAB stochastic reachability toolbox. In: Proceedings of the 22nd ACM International Conference on Hybrid Systems: Computation and Control, pp. 33–38 (2019)
Wang, Q., Zuliani, P., Kong, S., Gao, S., Clarke, E.M.: SReach: a probabilistic bounded deltareachability analyzer for stochastic hybrid systems. In: Roux, O., Bourdon, J. (eds.) CMSB 2015. LNCS, vol. 9308, pp. 15–27. Springer, Cham (2015). https://doi.org/10.1007/9783319234014_3
Acknowledgment
The authors would like to thank Thomas Gabler for his help in implementing traditional serial algorithms for the purpose of analysis and then comparing with the parallel ones.
Author information
Authors and Affiliations
Corresponding author
Editor information
Editors and Affiliations
Rights and permissions
Open Access This chapter is licensed under the terms of the Creative Commons Attribution 4.0 International License (http://creativecommons.org/licenses/by/4.0/), which permits use, sharing, adaptation, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons license and indicate if changes were made.
The images or other third party material in this chapter are included in the chapter's Creative Commons license, unless indicated otherwise in a credit line to the material. If material is not included in the chapter's Creative Commons license and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder.
Copyright information
© 2020 The Author(s)
About this paper
Cite this paper
Lavaei, A., Khaled, M., Soudjani, S., Zamani, M. (2020). AMYTISS: Parallelized Automated Controller Synthesis for LargeScale Stochastic Systems. In: Lahiri, S., Wang, C. (eds) Computer Aided Verification. CAV 2020. Lecture Notes in Computer Science(), vol 12225. Springer, Cham. https://doi.org/10.1007/9783030532918_24
Download citation
DOI: https://doi.org/10.1007/9783030532918_24
Published:
Publisher Name: Springer, Cham
Print ISBN: 9783030532901
Online ISBN: 9783030532918
eBook Packages: Computer ScienceComputer Science (R0)