AbstractA Spanish University has developed a novel system that integrates computational techniques concerning dynamics systems and intelligent learning schemes. The goal is to design efficient optimal control algorithms in order to be applied to any non-linear dynamic system. Technical cooperation agreements are sought in industrial environments and also for attitude control in satellites in space environment.
DetailsA Space Research Group (SRG) within a Spanish University specialised in different areas related to Aerospace (including Optimal Control), designs and builds instrumentation to be used on-board satellites, spacecrafts and unmanned aerial vehicles.
The research carried out by the SRG Group lets us have available an algorithm which is very robust against changes in the environment or in physic structure of the controlled system. For this reason, the learning takes into account of an implicit way, these possible changes. This algorithm is a closed-loop solution and both in the application of motion planning of mobile platforms (with movement forwards and backwards), as in the attitude control of satellites, learning is focused in reaching a specific objective from any original state, according to an optimization criterion (i.e. minimum time).
Thus, in this way, the controlled system (mobile platform or satellite) learns the dynamics and kinematics from the experience without any kind of explicit mathematical models. The new algorithm includes the ability and capacity to be executed either in two phases: learning and planning/optimal control; or in one stage of a concurrent way. The great advantage of the concurrency is to adapt the controller while it is being applied and, thus, achieving a behaviour that comes close to optimum.
The new algorithm acts on the controlled system through control actions in order to modify the state variables that characterise it. Therefore, the set of state variables and their values range constitute the so-called state space which will be the scope of the algorithm. Sometimes, in order to reduce learning times, it is important to take into account while it is learning possible symmetries among state variable and the possibility of extrapolating local knowledge to other areas of the state space.
Another unique feature of the algorithm is the absence of criticality in the choice of sampling period. Regardless of chosen value, controllability will not be affected because the period to be internally used by the algorithm is automatically adjusted to adjacency condition. This condition is based on the discretization of each of state variables. Therefore, each variable is divided into a number of cells with a specific size. Thus, a valid state change (also called transition) will occur when the distance between the initial state and the new one is at least the condition of adjacency established. This condition is measured in number of cells in any of the variables.
- No mathematical model of the controlled system. Instead, it is automatically generated in real time as the system learns from the experience.
- Independence in the choice of the sampling period of the system.
- Concurrency between learning and planning/control in closed-loop.
- Optimal behaviour with respect to an optimization criterion (minimum time, minimum energy or shortest path).