Abstract: The techniques disclosed herein mitigate stochasticity when controlling a mechanical system with artificial intelligence (AI) agents. In some configurations, AI agents are created using data generated by a machine learning model. Stochasticity is segmented temporally into near term and long term, and different strategies are used to address stochasticity in the different timeframes. For example, long term stochasticity may be addressed with changes to the reward function used to train the model. Short term stochasticity may be addressed by applying a margin to the output of an AI agent. Example margins include window averaging, clamps, and statistical process control bounds. In one configuration, AI agents are regression brains that are generated from setpoints inferred by the model from environmental states. The limitations inherent to fitting a regression line to this data may result in some predicted setpoints being outside of an allowed range.

Abstract: The techniques disclosed herein enable utilizing a full range of setpoint values to control a mechanical system. A machine learning model is trained with states collected from the mechanical system. Some of the states may have little to no variation, limiting exploration of possible setpoint values when training the model. To enable a more thorough exploration of possible setpoint values, the states are augmented with a fluctuating delta value that is derived from a fixed setpoint value. For example, a delta outside air temperature may be computed by subtracting outside air temperature, which fluctuates, from a fixed chilled water setpoint. A method of moments computation converts delta values inferred by the model back into absolute values. The absolute values are used to compute a regression equation that is usable by the mechanical system to compute a setpoint action for a given set of input states.