Algorithm consolidation
View Patent ↗This disclosure relates to a computer-implemented method of controlling an algorithm, and a device for implementing said method, the method including developing the algorithm from a first state to a second state by a machine learning process; determining a second computational cost of executing the algorithm in its second state; determining if the second computational cost satisfies a trigger condition; and, if so, consolidating the algorithm from the second state to a third state, wherein a third computational cost of executing the algorithm in its third state is less than the second computational cost.
1. A computer-implemented method of controlling an algorithm, the method comprising:
providing the algorithm in a first state, wherein the algorithm in each of a plurality of states of development is executable by a computing device using computational resources to produce one or more outputs based on one or more inputs, the algorithm in the first state has a first computational cost associated with computational resources for producing the one or more outputs, and the first computational cost is below a predetermined threshold;
developing the algorithm from the first state to a second state by a machine learning process, wherein the second state has a second computational cost associated with computational resources for producing the one or more outputs;
determining the second computational cost, the second computational cost being greater than the first computational cost;
determining that the second computational cost is above the predetermined threshold; and
based on the second computational cost being above the predetermined threshold, developing the algorithm from the second state to a third state by the machine learning process, wherein developing the algorithm from the second state to the third state comprises developing the algorithm such that the third state has a third computational cost associated with computational resources for producing the one or more outputs less than the predetermined threshold, wherein each of the first state, the second state, and the third state are different.
2. The method as claimed in claim 1 , further comprising:
executing the algorithm in the second state on a first set of inputs relating to a computing system so as to produce a second set of outputs, wherein each of the second set of outputs is one of a first possible output or a second possible output;
executing the algorithm in the third state on the first set of inputs so as to produce a third set of outputs, wherein each of the third set of outputs is one of the first possible output or the second possible output; and
determining a first matching value between the second set of outputs and the third set of outputs.
3. The method as claimed in claim 2 , further comprising determining whether the first matching value satisfies a new threshold.
4. The method as claimed in claim 3 , further comprising:
developing the algorithm from the third state to a fourth state,
wherein if the first matching value satisfies the new threshold then the algorithm is developed at a first learning rate, and if the first matching value does not satisfy the new threshold then the algorithm is developed at a second learning rate.
5. The method as claimed in claim 1 , wherein the algorithm is a decision tree having a plurality of branch nodes and a plurality of leaf nodes, and the computational cost is determined based on a count of the plurality of branch nodes.
6. The method as claimed in claim 1 , wherein developing the algorithm from the second state to the third state includes developing a new algorithm.
7. The method as claimed in claim 1 , wherein developing the algorithm from the second state to the third state includes modifying the algorithm from the second state.