-
Notifications
You must be signed in to change notification settings - Fork 0
Decision Process
A spider's data collection and the construction of a quickly executable model or function from that data do not factor into the moment-by-moment decision process of a spider. Instead, the decision process simply uses the constructed model to understand the likely outcomes of an action. In other words, when the spider needs to make a decision, the experiential model is a given; the decision process treats the model as a black box. The spiders's decision process then comes in interpreting what the model says.
In the current plans, the environmental features play no part in the solution space; these are given and cannot be modified by the spider. Instead, control features and the time axis comprise the solution space.
Even though the environmental features are viewed as constants, they change with every time-step. This means that the solution space constantly changes. This complicates the search for global maximums, but also dictates the solution space will only change as far as the environmental variables change. When change happens over time-steps, the turbulence of the solution space is limited over single time-steps.
Searching the whole space every time-step with random walks, examining the derivative for extrema, etc. are intractable given a sufficiently complex solution space. But because small changes in the environment cause similarly small changes in the solution space, searching the whole space is unnecessary. If a maximum was found in a the previous time-step, that same maximum very likely exists in this time-step, not far from where it was in the last time-step. Thus, the whole space needn't be searched; a simple and short burst of gradient ascent from the previous time-step's maximum location will very likely find this time-step's maximum.
The value of the sensor is not the only consideration in determining the best solution.