TimeMachineOptions<T>
Options & Configuration · AiDotNet.Models.Options
Configuration options for TimeMachine (Time Series State Space Model).
Inheritance: ModelOptions → RegressionOptions<T> → TimeSeriesRegressionOptions<T> → TimeMachineOptions<T>
For Beginners
TimeMachine is a modern architecture that combines ideas from state space models (like Mamba and S4) with a time-series-specific branch structure. RevIN first normalizes each series. E1 embeds the history length, two outer Mambas scan complementary orientations, E2 forms a smaller embedding for two inner Mambas, and P1/P2 combine residual and concatenated paths into the forecast.
Advantages:
- Linear complexity O(n) from SSM backbone
- Complementary scans model both channel and embedding-axis dependencies
- Residual paths preserve the original embedded signal
- State-of-the-art results on time series benchmarks
How It Works
TimeMachine is a state space model specifically designed for time series forecasting with two outer and two inner Mamba branches.
Reference: Ahamed et al., “TimeMachine: A Time Series is Worth 4 Mambas for Long-term Forecasting”, 2024.
https://arxiv.org/abs/2403.09898
MIGRATION — four properties were removed:NumScales, NumLayers,
UseMultiScaleAttention and TemporalDecompositionMethod. Each was public, settable
and read by nothing. The published graph is fixed — exactly four Mambas in the two-outer /
two-inner arrangement, combined by addition and concatenation rather than attention — so setting
any of them produced the identical model with no error and no signal. Delete the assignments;
there is no replacement, because there was never a behavior to replace. A build error naming the
property is the point: it tells you the setting was doing nothing, which the old code did not.
The dimensions that DO change the model — ModelDimension,
StateDimension, ExpandFactor, ConvKernelSize,
ContextLength, ForecastHorizon and DropoutRate —
are unchanged.
Constructors
| Constructor | Summary |
|---|---|
TimeMachineOptions() |
Initializes a new instance of the TimeMachineOptions class with default values. |
TimeMachineOptions(TimeMachineOptions<T> other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
bool UseReversibleNormalization { get; set; } |
Gets or sets whether to use reversible instance normalization. |
double DropoutRate { get; set; } |
Gets or sets the dropout rate for regularization. |
int ContextLength { get; set; } |
Gets or sets the context length (input sequence length). |
int ConvKernelSize { get; set; } |
Gets or sets the convolution kernel size for local context. |
int ExpandFactor { get; set; } |
Gets or sets the expansion factor for SSM inner dimension. |
int ForecastHorizon { get; set; } |
Gets or sets the forecast horizon (prediction length). |
int ModelDimension { get; set; } |
Gets or sets the model dimension (d_model). |
int StateDimension { get; set; } |
Gets or sets the state dimension for each SSM block. |