Determining causal models for controlling environments
Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for determining causal models for controlling environments. One of the methods includes obtaining data specifying baseline probability distributions for each of a plurality of controllable elements; maintaining a causal model; repeatedly performing the following: selecting control settings for the environment based on the causal model and values for a particular internal parameter of the control system that are sampled from a range of possible values; selecting control settings for the environment based on the baseline probability distributions; monitoring environment responses to the control settings selected based on the causal model and the control settings selected based on the baseline probability distributions; determining, for each of the possible values, a measure of a difference between a current system performance and a baseline system performance; and updating how frequently each of the possible values is sampled.
1. A method for controlling an environment, the method comprising:
obtaining data specifying baseline probability distributions over possible settings for each of a plurality of controllable elements of an environment;
maintaining a causal model that identifies causal relationships between the possible settings and environment responses to the possible settings;
repeatedly performing the following:
selecting, by a control system, control settings for the environment based on the causal model and values for a particular internal parameter of the control system that are sampled from a range of possible values for the particular internal parameter;
selecting control settings for the environment based on the baseline probability distributions;
monitoring environment responses to (i) the control settings selected based on the causal model and (ii) the control settings selected based on the baseline probability distributions;
determining, for each of the possible values of the particular internal parameter, a measure of a difference between (i) a current system performance when the control settings are selected based on the causal model and on the possible value of the particular internal parameter and (ii) a baseline system performance measure for the control settings selected based on the baseline probability distributions; and
updating how frequently each of the possible values is sampled based on the measures of difference.
2. The method of claim 1 , wherein the particular parameter is a data inclusion window that defines how many previously received environment responses are used to update the causal model.
3. The method of claim 1 , wherein the particular parameter is a temporal extent parameter that defines which environment responses are associated with which selection of control settings in the causal model.
4. The method of claim 1 , wherein the particular parameter is a ratio parameter that defines how frequently control settings are selected using the baseline probability distributions relative to being selected using the causal model.
5. The method of claim 1 , wherein the particular parameter is a clustering parameter that is a hyperparameter of a clustering technique employed to cluster procedural instances in the environment into clusters when selecting control settings for the procedural instances.
6. The method of claim 1 , the method further comprising:
maintaining a second causal model that identifies causal relationships between the possible values for the particular internal parameter and the measure of the difference, wherein:
the possible values are sampled based on the second causal model, and
updating how frequently each of the possible values comprises updating the second causal model based on the measures of difference.
7. The method of claim 6 , wherein the second causal model comprises, for each possible value for the particular internal parameter, (i) an impact measurement that measures the impact of the possible value on the measure of difference and (ii) a confidence interval that represents a level of uncertainty about the impact measurement.
8. The method of claim 6 , wherein the possible values are sampled by mapping the impact measurements and the confidence intervals to a probability distribution over the possible values and sampling a value from the probability distribution.
9. The method of claim 5 , further comprising:
adjusting the range of possible values for the particular internal parameter during operation of the control system.
10. The method of claim 9 , further comprising:
adjusting the range of possible values based on a statistical power analysis of a statistical power of the second causal model.
11. A system comprising one or more computers and one or more storage devices storing instructions that when executed by the one or more computers cause the one or more computers to perform the operations of the method of claim 1 .
12. One or more computer-readable storage media storing instructions that when executed by one or more computers cause the one or more computers to perform the operations of the method of claim 1 .