swifttd 0.1.9__cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl

swifttd/__init__.py

@@ -0,0 +1,5 @@
+from ._version import __version__
+# Import C++ bindings module built by pybind11_add_module (swift_td)
+from swift_td import SwiftTDNonSparse, SwiftTDBinaryFeatures, SwiftTD
+
+__all__ = ["SwiftTDNonSparse", "SwiftTDBinaryFeatures", "SwiftTD", "__version__"]

swifttd/_version.py

@@ -0,0 +1 @@
+__version__ = "0.1.9"

swifttd-0.1.9.dist-info/METADATA

@@ -0,0 +1,93 @@
+Metadata-Version: 2.2
+Name: swifttd
+Version: 0.1.9
+Summary: SwiftTD: Fast and Robust TD Learning
+Author: Khurram Javed
+License: MIT
+Classifier: Programming Language :: Python :: 3
+Classifier: Programming Language :: C++
+Classifier: Operating System :: OS Independent
+Project-URL: Homepage, https://github.com/khurramjaved96/SwiftTD
+Requires-Python: >=3.7
+Description-Content-Type: text/markdown
+
+# SwiftTD: A Fast and Robust Algorithm for Temporal Difference Learning
+
+SwiftTD is an algorithm for learning value functions. It combines the ideas of step-size adaptation with the idea of a bound on the rate of learning. The implementations in this repository use linear function approximation. 
+
+## Installation
+
+```bash
+pip install SwiftTD
+```
+
+## Usage
+
+After installation, you can use the three implementations of SwiftTD in Python as: 
+
+```python
+import swifttd
+
+# Version of SwiftTD that expects the full feature vector as input. This should only be used if the feature representation is not sparse. Otherwise, the sparse versions are more efficient.
+td_dense = swifttd.SwiftTDNonSparse(
+    num_of_features=5,     # Number of input features
+    lambda_=0.95,        # Lambda parameter for eligibility traces
+    alpha=1e-2,  # Initial learning rate
+    gamma=0.99,        # Discount factor
+    epsilon=1e-5,          # Small constant for numerical stability
+    eta=0.1, # Maximum allowed step size (bound on rate of learning)
+    decay=0.999, # Step size decay rate
+    meta_step_size=1e-3,  # Meta learning rate
+    eta_min=1e-10 # Minimum value of the step-size parameter
+)
+
+# Feature vector
+features = [1.0, 0.0, 0.5, 0.2, 0.0] 
+reward = 1.0
+prediction = td_dense.step(features, reward)
+print("Dense prediction:", prediction)
+
+# Version of SwiftTD that expects the feature indices as input. This version assumes that the features are binary---0 or 1. For learning, the indices of the features that are 1 are provided. 
+td_sparse = swifttd.SwiftTDBinaryFeatures(
+    num_of_features=1000,     # Number of input features
+    lambda_=0.95,        # Lambda parameter for eligibility traces
+    alpha=1e-2,  # Initial learning rate
+    gamma=0.99,        # Discount factor
+    epsilon=1e-5,          # Small constant for numerical stability
+    eta=0.1, # Maximum allowed step size (bound on rate of learning)
+    decay=0.999, # Step size decay rate
+    meta_step_size=1e-3,  # Meta learning rate
+    eta_min=1e-10 # Minimum value of the step-size parameter
+)
+
+# Specify the indices of the features that are 1.
+active_features = [1, 42, 999]  # Indices of active features
+reward = 1.0
+prediction = td_sparse.step(active_features, reward)
+print("Sparse binary prediction:", prediction)
+
+# Version of SwiftTD that expects the feature indices and values as input. This version does not assume that the features are binary. For learning, it expects a list of (index, value) pairs. Only the indices of the features that are non-zero need to be provided. 
+
+td_sparse_nonbinary = swifttd.SwiftTD(
+    num_of_features=1000,     # Number of input features
+    lambda_=0.95,        # Lambda parameter for eligibility traces
+    alpha=1e-2,  # Initial learning rate
+    gamma=0.99,        # Discount factor
+    epsilon=1e-5,          # Small constant for numerical stability
+    eta=0.1, # Maximum allowed step size (bound on rate of learning)
+    decay=0.999, # Step size decay rate
+    meta_step_size=1e-3,  # Meta learning rate
+    eta_min=1e-10 # Minimum value of the step-size parameter
+)
+
+# Specify the indices and values of the features that are non-zero.
+feature_values = [(1, 0.8), (42, 0.3), (999, 1.2)]  # (index, value) pairs
+reward = 1.0
+prediction = td_sparse_nonbinary.step(feature_values, reward)
+print("Sparse non-binary prediction:", prediction)
+```
+
+## Resources
+- [Paper (PDF)](https://khurramjaved.com/swifttd.pdf)
+- [Interactive Demo](https://khurramjaved.com/swifttd.html)
+

swifttd-0.1.9.dist-info/RECORD

@@ -0,0 +1,6 @@
+swift_td.cpython-310-aarch64-linux-gnu.so,sha256=4rJ0QwyV6btRWOZ8EEYzdaNpATGUIpt6Qnp-ncm1m90,266744
+swifttd/__init__.py,sha256=IiG3Lrza6VD39Z7mePN1DUl_ND_hsqLzIvHa1WRdSRA,256
+swifttd/_version.py,sha256=XIaxbMbyiP-L3kguR1GhxirFblTXiHR1lMfDVITvHUI,22
+swifttd-0.1.9.dist-info/METADATA,sha256=D1NCKI2Z2X5j0MV5p8gyqX-RswEZWYdUlb1Eu4sywAU,3890
+swifttd-0.1.9.dist-info/WHEEL,sha256=e8r2Onm5NNvlP7oexPErjEYtutEfns1NIR2WhYHCUPg,158
+swifttd-0.1.9.dist-info/RECORD,,

swifttd-0.1.9.dist-info/WHEEL

@@ -0,0 +1,6 @@
+Wheel-Version: 1.0
+Generator: scikit-build-core 0.11.6
+Root-Is-Purelib: false
+Tag: cp310-cp310-manylinux_2_17_aarch64
+Tag: cp310-cp310-manylinux2014_aarch64
+