Milestone 3 sanjosespartans DarkHorseIdea: Task Recommendation and Evaluation using Reinforcement Learning or Evolutionary Algorithms

From crowdresearch
Jump to: navigation, search

Using Reinforcement Learning:

An reinforcement learning technique can be adopted where the model learns task recommendation and task evaluation based on rewards and punishment. The model can be rewarded for a relevant task recommendation (e.g. when the worker picks a method from the tasks it suggested) or penalised for a bad recommendation. It can be expected that over the time the model will learn appropriate weights to appropriate categories.


Using Evolutionary Algorithm:

An evolutionary algorithm can be deployed, where the initial population is say a collection of task_recommendations (randomly generated).

Each of these task_recommendations can be represented in form of a chromosome representing, for example, the task recommendation category and skills associated.

Then two of these chromosomes can be selected as parents using roulette selection or other mechanism.

The selected parents can be crossed at random points or using some crossing technique to give child chromosome representing recommendation for combination of skills. If the child chromosome is superior to the parents it can be added to the population pool in lieu of the parents to become more fit. Mutation can be introduced with a small probabality in order to maintain diversity.

The process can continue for certain number of iterations or till a specific criteria is met. It can be expected that the population will ultimately converge to have reasonable recommendations given a task.

The same technique can be adopted for model evaluation to ultimately give a model that can evaluate the quality of work produced so that workers can be rated fairly and requesters can trust the rating more.