All issues

In this paper, we propose an adaptive traffic signal control method invariant to the configuration of the traffic signal. The proposed method uses one neural network model to control traffic signals of various configurations, differing both in the number of controlled lanes and in the used traffic light control cycle (set of phases). To describe the state space, both dynamic information about the current state of the traffic flow and static data about the configuration of a controlled intersection are used. To increase the speed of model training and reduce the required amount of data required for model convergence, it is proposed to use an “expert” who provides additional data for model training. As an expert, we propose to use an adaptive control method based on maximizing the weighted flow of vehicles through an intersection. Experimental studies of the effectiveness of the developed method were carried out in a microscopic simulation software package. The obtained results confirmed the effectiveness of the proposed method in different simulation scenarios. The possibility of using the developed method in a simulation scenario that is not used in the training process was shown. We provide a comparison of the proposed method with other baseline solutions, including the method used as an “expert”. In most scenarios, the developed method showed the best results by average travel time and average waiting time criteria. The advantage over the method used as an expert, depending on the scenario under study, ranged from 2% to 12% according to the criterion of average vehicle waiting time and from 1% to 7% according to the criterion of average travel time.

Distributed saddle point problems (SPPs) have numerous applications in optimization, matrix games and machine learning. For example, the training of generated adversarial networks is represented as a min-max optimization problem, and training regularized linear models can be reformulated as an SPP as well. This paper studies distributed nonsmooth SPPs with Lipschitz-continuous objective functions. The objective function is represented as a sum of several components that are distributed between groups of computational nodes. The nodes, or agents, exchange information through some communication network that may be centralized or decentralized. A centralized network has a universal information aggregator (a server, or master node) that directly communicates to each of the agents and therefore can coordinate the optimization process. In a decentralized network, all the nodes are equal, the server node is not present, and each agent only communicates to its immediate neighbors.

We assume that each of the nodes locally holds its objective and can compute its value at given points, i. e. has access to zero-order oracle. Zero-order information is used when the gradient of the function is costly, not possible to compute or when the function is not differentiable. For example, in reinforcement learning one needs to generate a trajectory to evaluate the current policy. This policy evaluation process can be interpreted as the computation of the function value. We propose an approach that uses a smoothing technique, i. e., applies a first-order method to the smoothed version of the initial function. It can be shown that the stochastic gradient of the smoothed function can be viewed as a random two-point gradient approximation of the initial function. Smoothing approaches have been studied for distributed zero-order minimization, and our paper generalizes the smoothing technique on SPPs.

The problem is to automate the correction of LaTeX documents. Each document is represented as a parse tree. The modified Zhang-Shasha algorithm is used to construct a mapping of tree vertices of the original document to the tree vertices of the edited document, which corresponds to the minimum editing distance. Vertex to vertex maps form the training set, which is used to generate rules for automatic correction. The statistics of the applicability to the edited documents is collected for each rule. It is used for quality assessment and improvement of the rules.

High frequency algorithmic trading became is a subclass of trading which is focused on gaining basis-point like profitability on sub-second time frames. Such trading strategies do not depend on most of the factors eligible for the longer-term trading and require specific approach. There were many attempts to utilize machine learning techniques to both high and low frequency trading. However, it is still having limited application in the real world trading due to high exposure to overfitting, requirements for rapid adaptation to new market regimes and overall instability of the results. We conducted a comprehensive research on combination of known quantitative theory and reinforcement learning methods in order derive more effective and robust approach at construction of automated trading system in an attempt to create a support for a known algorithmic trading techniques. Using classical price behavior theories as well as modern application cases in sub-millisecond trading, we utilized the Reinforcement Learning models in order to improve quality of the algorithms. As a result, we derived a robust model which utilize Deep Reinforcement learning in order to optimise static market making trading algorithms’ parameters capable of online learning on live data. More specifically, we explored the system in the derivatives cryptocurrency market which mostly not dependent on external factors in short terms. Our research was implemented in high-frequency environment and the final models showed capability to operate within accepted high-frequency trading time-frames. We compared various combinations of Deep Reinforcement Learning approaches and the classic algorithms and evaluated robustness and effectiveness of improvements for each combination.