Self-Learning Algorithms
Simple statistical approaches cannot adequately reveal the learning process. Within an LMS, tests are rarely designed to repeat the tasks at an identical level. Therefore, measurement models that are based on the classical test theory usually estimate the learning success within a criterion only inadequately. They only offer an extra perspective to better understand the exact learning process within a criterion. However, the are not appropriate for an overview.
In order to conceive the differences between tasks and persons in one, different test theories have been developed. The best known is the item-response theory which include Rasch models. However, they place highly formal demands on the data material, which cannot be fulfilled in an LMS without special precautions. Alternatively, the LMSA Kit uses self-learning algorithms in the field of matchmaking. They can be used for educational purposes and to put much lower demands on the data material.
Within the LMSA Kit different matchmaking procedures are implemented.
ELO
The ELO algorithm is one of the first matchmaking algorithms. It comes from online chess. It tries to assign each player a skill from the results of already played chess games, so that future matches with equal opponents can be set. The ELO algorithm has remained in the chess world up to now. Even today, the achievements of chess players are assessed by their ELO number and the title Grand Master may be worn, for example, only if (among other things) at least once an ELO number of over 2500 was reached. The LMSA Kit can use this algorithm to evaluate a performance within a criterion.
ELO K
The actual ELO algorithm includes a term that models its learning behavior. This term usually minimizes the rate of change in the future, so that the score initially settles quickly to an initial level and later only increases with stable improvements. In the variant ELO K, the two threshold values of the ELO algorithm are no longer taken from the original chess variant, but can be adapted by the user.
ELO UN
Instead of specifying fixed k-values, a continuous rational modeling of this learning behavior is attempted here, depending on the number of tasks.
k = U / (1+ N * solved tasks) with U, N rational numbers.
Glicko Version 1
The Glicko system was developed by Mark E. Glickman. It uses a similar procedure as the ELO algorithm. The change of the ability values, if there is a new result, however, no longer takes place via a rigidly fixed k, but is determined on a case-by-case basis, depending on the ability of both counterparties. Unexpected results lead to stronger changes. If a low-skill learner can solve a difficult task, his/her skill value will increase more than the one of learners with a high skill value.
Various parameters can be set here as parameters, which are used primarily for calculating this rate of change. A more detailed description will follow shortly.
TrueSkill
TrueSkill is a matchmaking process developed by Microsoft. It is, for instance, implemented on the XBox Live platform in order to set exciting matches here. It uses various Bayesian estimation techniques, which are utilized in a kind of factor graph to deduce the actual capability behind the shown performance.
The plug-in code to implement the TrueSkill algorithm was originally written by Jeff Moser <jeff@moserware.com>. The exact license can be found there in the GitHub repository or in the plug-in information dialog within the LMSA Kit.