Reinforcement learning

Our research project in the field of Reinforcement learning: Actor-critic Deep-Q recurrent networks for temporal sequencing

Project description:

Starting from developing new techniques of time series analysis, through the use of virtual environments for the training of model-free agents, we want to generate an actor/critic agent (Barto et al.1988) that optimizes and surpasses the performances of forecast models currently available.


  • Develop mathematical models that provide plausible descriptions starting from sample data.
  • Understand or model the stochastic mechanisms that give rise to an observed series and predict future values.
  • Creating a model that best fits available historical data and uses it for future observations.

Possible applications:

  • Yearly forecasting of the yield of the corn in tons per state.
  • Predict if an EEG trace in seconds indicates that a patient has an attack or not.
  • Daily forecasting of the stock closing price.
  • Yearly Forecasting of the birth rate in all the hospitals of a city every year.
  • Daily products sales prediction for a store.
  • Daily passengers prediction for a train station.
  • Forecast of unemployment for a state every quarter.
  • Hourly forecast application usage on one server.
  • Forecasting the size of the rabbit population in one state each breeding season.
  • Daily forecast of the average petrol price in a city.

Model generation using the latest languages ​​and machine learning libraries currently available.

(Tensorflow) Algorithm development using existing virtual environments (OpenAI Gym) or created ad hoc.

Creation of the hardware infrastructure necessary for the development, training and use of the model.