Simulation and automated control of physical systems

A computer system to simulate and control a physical system, the computer system comprising:

• a memory with instructions stored thereon;
• and
• a processing device, coupled to the memory, the processing device configured to access the memory and execute the instructions, wherein the instructions cause the processing device to implement a model-based reinforced-learning (MBRL) agent that perform operations comprising:
• forming a computational graph with a plurality of nodes initialized with respective parameters, wherein a structure of the computational graph is based on one or more physical processes in the physical system, wherein the forming includes defining a respective differential equation that specifies a time derivative for each of x and y;
• receiving, from one or more sensors, measured values of one or more observed states of the physical system;
• setting initial values of one or more unobserved states of the physical system, wherein the unobserved states each correspond to a physical state of the physical system that is not measured
• receiving values of one or more control inputs a for the physical system;
• and
• iteratively simulating the physical system using x, y and a as simulation inputs to the computational graph, wherein the simulating includes:
• solving the differential equations using an ordinary differential equation (ODE) solver to obtain a predicted value for x and y at a next time point;
• and
• adjusting at least one of the parameters of the computational graph based on a comparison of the predicted value of x at the next point and a groundtruth value of x at the next point as measured in the physical system.

An automated greenhouse, comprising:

• a foundation with at least one growth medium for growing one or more plants within an interior of the automated greenhouse;
• one or more controllers configured to alter one or more physical characteristics within an interior of the automated greenhouse;
• and
• a computer system configured to control the one or more controllers, the computer system comprising:
• a memory with instructions stored thereon;
• and
• a processing device, coupled to the memory, the processing device configured to access the memory and execute the instructions, wherein the instructions cause the processing device to implement a model-based reinforced-learning (MBRL) agent that perform operations comprising:
• forming a computational graph with a plurality of nodes initialized with respective parameters, wherein a structure of the computational graph is based on one or more physical processes in the automated greenhouse, wherein the forming includes defining a respective differential equation that specifies a time derivative for each of x, y, and z;
• receiving, from one or more sensors, measured values of one or more observed states of the automated greenhouse;
• setting initial values of:
• one or more unobserved states of the automated greenhouse, wherein the unobserved states each correspond to a physical state of the automated greenhouse that is not measured;
• and
• one or more latent states for the automated greenhouse, wherein the latent states correspond to one or more additional factors that affect the automated greenhouse;
• receiving values of one or more control inputs a for the automated greenhouse;
• and
• iteratively simulating the automated greenhouse using x, y, z, and a as simulation inputs to the computational graph, wherein the simulating includes:
• solving the differential equations using an ordinary differential equation (ODE) solver to obtain a predicted value for x, y, and z at a next time point;
• and
• adjusting at least one of the parameters of the computational graph based on a comparison of the predicted value of x at the next point and a groundtruth value of x at the next point as measured in the automated greenhouse.