SuperMemo 17 uses a new spaced repetition algorithm denoted Algorithm SM-17. Unlike all prior algorithms that were either theoretical or "inspired by data", this algorithm has been developed entirely on the basis of prior records of repetitions collected by users of SuperMemo. This data-driven effort required untold hours of analysis while processing millions of repetition samples. Tools : Memory : 4D Graphs was instrumental in that analysis and debugging process. If you want to understand the algorithm and help improve it further, please study those tools and keep analyzing your own data and your own memory. In a stochastic system of memory, perfection is impossible, but we should always try to come closer to the optimum.
All memory graphs provide a 3-dimensional view with rotation along all 3 axes (X, Y and Z), and a slider for animation in the 4th dimension along item difficulty.
Figure: 3D graph of SInc matrix based on 60,167 repetition cases for items with difficulty=0.5. The increase is dramatic at low stabilities (17-fold increase is unheard of in earlier SuperMemos), and peaks at retrievability of 0.85. In some cases, the SInc drops below 1.0, which corresponds with a drop in stability (i.e. memory lability). Those low values of SInc do not depend on retrievability. Those huge variations in SInc are the main reason why SuperMemo 17 beats SuperMemo 16 in learning metrics by a wide margin..
Figure: A "from above" view at the SInc matrix providing a contour map. Red zones indicate high stability increase at review. The picture shows that the greatest stability increase occurs for lower stability levels and retrievabilities around 70-90%.
Figure: Approximating the SInc matrix with the best-fit function used by default in SuperMemo to compute the increase in stability (e.g. in cases of lack of data). The approximation procedure uses a hill-climbing algorithm with parameters A, B, C, D displayed in the picture. Least squares deviation is obtained to asses the progress. Green circles represent the Sinc matrix at a chosen difficulty level. Their size corresponds with repetition cases investigated. The blue surface is the best fit of the studied function to the SInc data.
Figure: The Recall matrix graph shows that the actual recall differs from predicted retrievability. For higher stabilities and difficulties, it is harder to reach the desired recall level.
Figure: Approximating the Recall matrix with the best-fit function to compute default recall in conditions of data scarcity. The approximation procedure uses a hill-climbing algorithm with parameters A, B, C, D displayed in the picture. Least squares deviation is obtained to asses the progress. The circles represent the Recall matrix at a chosen difficulty level. Their size corresponds with repetition cases investigated. The red surface is the best fit of the studied function to the Recall data.
Figure: Approximating the Recall matrix with the best-fit function to compute default recall in conditions of data scarcity. By choosing the right viewing angle, the curve that reflects the changes to recall with retrievability can be seen in abstraction of stability. In this case the relationship is almost linear (the logarithmic bend is a result of the log scale used for retrievability).
Figure: The relationship between the first interval after failure, retrievability at review, and prior memory lapse count.