Method for converting dynamical systems with continuous states into Markov decision processes with discrete states
View Patent ↗A continuous dynamical system is converted to a Markov decision process (MDP) with discrete states. A predetermined number of continuous states of the continuous system is selected, wherein each continuous state corresponds to one discrete state of the MDP. Delaunay triangulation is applied to the continuous states to produce a set of triangles, wherein vertices of each triangle represent the continuous states. For each discrete state, a next discrete state y=ƒ(x, a) is determined, wherein x represents the continuous state corresponding to the discrete state, a is a control action, and ƒ is a non-linear transition function for the continuous. A particular triangle containing the next discrete state y is identified, and the next discrete state y is expressed as probabilities of transitioning to the discrete states corresponding to the continuous states x represented by the vertices of the particular triangle.
1. A method for converting a continuous system to a Markov decision process (MDP), wherein the continuous system is dynamical and the MDP has discrete states, comprising the steps of:
selecting a predetermined number of continuous states of the continuous system, wherein each continuous state corresponds to one discrete state of the MDP; and
applying Delaunay triangulation to the continuous states to produce a set of triangles, wherein vertices of each triangle represent the continuous states, and further comprising for each discrete state;
determining a next discrete state y=ƒ(x,a), wherein x represents the continuous state corresponding to the discrete state, a is a control action, and ƒ is a non-linear transition function for the continuous;
identifying a particular triangle containing the next discrete state y; and
expressing the next discrete state y as probabilities of transitioning to the discrete states corresponding to the continuous states x represented by the vertices of the particular triangle, where identifying the particular triangle is performed by means of forming a d×d difference matrix E, such that column j of the difference matrix E contains a differences v m,j −v m.d+1 , for j=1, d, between an j th vertex v m.j of triangle m and a d+1 st vertex q=v m.d+1 of the triangle m, and solving a linear system of equations Ec=(y−q), for a vector c, and wherein the triangle m is identified as containing y if all c j >0, j=1, d, and c d+1 =1−Σ j=1 D c j >0, wherein the steps are performed in a processor.
2. The method of claim 1 , where the discrete states of the MDP are selected by uniform random sampling of a state space X of the continuous system.
3. The method of claim 1 , where the discrete states of the MDP are selected by imposing a regular grid on a state space X of the continuous system.
4. The method of claim 1 , where an inverse matrix E −1 of the matrix E is pre-computed and stored, such that the linear equation can be solved as c =E −1 (y−q).
5. The method of claim 1 , where the triangles in the Delaunay triangulation are traversed in increasing order of a distance from centroids of the vertices to the state y.
6. The method of claim 1 , wherein the triangles are generalized to simplices.