Tata Consultancy Services Limited

Avinash Achar

Venkatesh Sarangan

Anand Sivasubramaniam

Easwara Naga Subramanian

Sanjay Purushottam Bhat

Soumen Pachal

Yogesh Kumar Bichpuriya

Abhay Pratap Singh

Akshaya Natarajan

Sanjay Krupakar Bhat

Arunchandar Vasan

Rajesh Jayaprakash

Srinarayana Nagarathinam

Balaraman Ravindran

Gokul Karthik

Rohith Regikumar

Vignesh Lakshmanan Kangadharan Palani Radja

Antony Joshini Lobo

Vignesh Lakshmanan Kangadharan Palaniradja

Nancy Bhutani

Arrival/Travel times for public transit exhibit variability on account of factors like seasonality, dwell times at bus stops, traffic signals, travel demand fluctuation, spatial and temporal correlations, etc. The developing world in particular is plagued by additional factors like lack of lane discipline, excess vehicles, diverse modes of transport and so on. This renders the bus arrival time prediction (BATP) to be a challenging problem especially in the developing world. Present disclosure provides system and method that implement recurrent neural networks (RNNs) for BATP (in real-time), wherein the system incorporates information pertaining to spatial and temporal correlations and seasonal correlations. More specifically, a Gated Recurrent Unit (GRU) based Encoder-Decoder (ED) model with one or more bi-directional layers at the decoder is implemented for BATP based on relevant additional synchronized inputs (from previous trips) at each step of the decoder. The system further captures congestion influences on travel time prediction.

Systems and methods for public transit arrival time prediction

Reinforcement Learning agent interacting with a real-world building to determine optimal policy may not be viable due to comfort constraints. Embodiments of the present disclosure provide multi-deep agent RL for dynamically controlling electrical equipment in buildings, wherein a simulation model is generated using design specification of (i) controllable electrical equipment (or subsystem) and (ii) building. Each RL agent is trained using simulation model and deployed in the subsystem. Reward function for each subsystem includes some portion of reward from other subsystem(s). Based on reward function of each RL agent, each RL agent learns an optimal control parameter during execution of RL agent in subsystem. Further, a global optimal control parameter list is generated using the optimal control parameter. The control parameters in the global optimal control parameters list are fine-tuned to improve subsystem's performance. Information on fine-tuning parameters of the subsystem and reward function are used for training RL agents.

Multi-agent deep reinforcement learning for dynamically controlling electrical equipment in buildings

This disclosure relates to method and system for optimal policy learning and recommendation for distribution task using deep RL model, in applications where when the action space has a probability simplex structure. The method includes training a RL agent by defining a policy network for learning the optimal policy using a policy gradient (PG) method, where the policy network comprising an artificial neural network (ANN) with a set of outputs. A continuous action space having a continuous probability simplex structure is defined. The learning of the optimal policy is updated based on one of stochastic and deterministic PG. For stochastic PG, a Dirichlet distribution based stochastic policy parameterized by output of the ANN with an activation function at an output layer of the ANN is selected. For deterministic PG, a soft-max function is selected as activation function at the output layer of the ANN to maintain the probability simplex structure.

Optimal policy learning and recommendation for distribution task using deep reinforcement learning model

Currently available time-series prediction techniques only factors last observed value from left of missing values and immediate observed value from right is mostly ignored while performing data imputation, thus causing errors in imputation and learning. Present application provides methods and systems for time-series prediction under missing data scenarios. The system first determines missing data values in time-series data. Thereafter, system identifies left data value, right data value, left gap length, right gap length and mean value for each missing data value. Further, system provides left gap length and right gap length identified for each missing data value to feed-forward neural network to obtain importance of left data value, right data value and mean value. The system then passes importance obtained for each missing data value to SoftMax layer to obtain probability distribution that is further utilized to calculate new data value corresponding to each missing data value in time-series data.

Methods and systems for time-series prediction under missing data using joint impute and learn technique

State of the art systems that are used for time series data prediction have the disadvantage that perform only one step prediction, which has only limited application. Disadvantage of such systems is that extent of applications of such single step predictions are limited. The disclosure herein generally relates to time series data prediction, and, more particularly, to time series data prediction based on seasonal lags. The system processes collected input data and determines order of seasonality of the input data. The system further selects encoders based on the determined order of seasonality and generates input data for a decoder that forms encoder-decoder pair with each of the encoders. The system then generates time series data predictions based on seasonal lag information distributed without redundance between encoder and decoder inputs.

Method and system for time series data prediction based on seasonal lags

This disclosure relates generally to a method and system for dynamically predicting vehicle arrival time using a temporal difference learning technique. Due to varying uncertainties predicting vehicle arrival time and travel time are crucial elements to make the public transport travel more attractive and reliable with increased traffic volumes. The method includes receiving a plurality of inputs in real time and then extracting a plurality of temporal events from a closest candidate trip pattern using a historical database. Further, a trained temporal difference predictor model (TTDPM) is utilized for dynamically predicting the arrival time from the current location of the vehicle to the target destination based on the plurality of nonlinear features. The non-linear features and linear approximator formulation of TTDPM provides fast gradient computation improves training time. Additionally, updating the revised state information at every iteration provides better accuracy of arrival time prediction in real time.

Method and system for dynamically predicting vehicle arrival time using a temporal difference learning technique

Sum of bid quantities (across price bands) placed by generators in energy markets have been observed to be either constant OR varying over a few finite values. Several researches have used simulated data to investigate desired aspect. However, these approaches have not been accurate in prediction. Embodiments of the present disclosure identified two sets of generators which needed specialized methods for regression (i) generators whose total bid quantity (TBQ) was constant (ii) generators whose total bid quantity varied over a few finite values only. In first category, present disclosure used a softmax output based ANN regressor to capture constant total bid quantity nature of targets and a loss function while training to capture error most meaningfully. For second category, system predicts total bid quantity (TBQ) of a generator and then predicts to allocate TBQ predicted across the various price bands which is accomplished by the softmax regression for constant TBQs.

Neural network based prediction of competition behaviour in energy markets

A system and method for predicting travel time of a vehicle on routes of unbounded length within arterial roads. It collects historical information from probe vehicles positions using GPS technology in a periodic fashion and the sequence of links traversed between successive position measurements. Further, it collects information of neighborhood structure for each link within the arterial roads network. Any of the existing conditional probability distribution functions could be used to capture the spatio-temporal dependencies between each link of the arterial network and its neighbors. It learns the parameters of this data driven probabilistic model from historical information of probe vehicle trajectories traversed within the arterial roads network using an associated expectation maximization method. Finally it predicts travel time of vehicles on routes of unbounded length in a novel fashion within the network of arterial roads using the learnt parameters and current real time information of trajectories of vehicle.

Predicting vehicle travel time on routes of unbounded length in arterial roads

In energy markets in which bidding process is used to sell energy, it is important that a mechanism for deciding bidding amount is in place. State of the art systems in this domain have the disadvantage that they rely on simulation data, and also they make certain assumptions, and both the factors can affect accuracy of results when the systems are deployed and are expected to handle practical scenarios. The disclosure herein generally relates to energy markets, and, more particularly, to a method and a system for Reinforcement Learning (RL) based model for generating bids. The system trains a RL agent using historical data with respect to competitor bids places and Market Clearing Prices (MCPs). The RL agent then processes real-time inputs and generates bidding recommendations.

Method and system for building reinforcement learning (RL) based model for generating bids

Accurate estimation of the trajectory of a vehicle by selecting optimal number of GPS data points and a shortest path technique applied for estimation is important and crucial. Method and system for estimating a trajectory from GPS data points is described. The method disclosed utilizes a plurality of GPS data points of a vehicle, an existing road map and a set of equal time intervals obtained by dividing an elapsed time during movement of the vehicle. Each GPS data point is associated to a time interval and a set of candidate points are mapped to each GPS data point correspondingly. A set of possible paths are determined between the set of candidate points in each time interval to estimate the trajectory of the vehicle using one of a shortest path technique and an edit distance technique.

Method and system for estimating a trajectory from GPS data points

This disclosure relates generally to method and system for forecasting sales based on N-Gram model. The present disclosure provides accurate prediction of sales for optimal operations to reduce the cost. The method receives a plurality of inputs of each product comprising a sales history, and a current price bin. The categorical sale(s) for each product is discretized based on the sales history by clustering each product sales history into a one or more groups based on a maximum sales velocity range. Further, a probability table is generated for the discretized categorical sales of each product based on computing a round off weighted mean and a median using a N-Gram model. Then, a smooth probability table is computed for the generated probability table. To forecast sales multistep prediction for the smooth probability table is computed based on at least one of a joint approach, a bootstrapped approach, and a step greedy approach.

Method and system for forecasting sales based on N-Gram model

A system and method for predicting travel time of a vehicle on one or more routes within arterial roads. It collects historical information from probe vehicles positions using GPS technology in a periodic fashion and the sequence of links traversed between successive position measurements. Further, it collects information of neighborhood structure for each link within the arterial roads network. A NoisyOR conditional probability distribution function is proposed to capture the spatio-temporal dependencies between each link of the arterial network and its neighbors. It learns the parameters of this data driven probabilistic model from collected historical information of probe vehicle trajectories traversed within the arterial roads network using a proposed expectation maximization method. Finally it predicts travel time of vehicles on routes within the arterial roads using the learnt parameters and real time information of trajectories of vehicle that have been recorded from the arterial roads network using GPS technology sensing.

Growing community of inventors

Chennai, India

Avinash Achar