• benchmark Includes comparative experiments with previous work.
• includeinclude Auxiliary.
• index Core components of Chameleon. Workspace is used for training Chameleon.
• others Baselines.

• C++17.
• NVIDIA GPU.
• PyTorch with the PyTorch C++ frontend (libtorch).

• Ubuntu 22.04.3 LTS
• AMD 7900X CPU @ 4.7GHz
• 128 GB DDR5 5200Mhz main memory
• NVIDIA GeForce RTX 4070 12GB
• CUDA 12.1
• PyTorch 2.1.0

• data_father_path path for storing datasets
• model_father_path path for storing models of RL
• experience_father_path path for storing experiments of DARE

• The example can be seen in any file that includes "CHA:: Index" in the benchmark test.
• To successfully run the code, you need to add the source code baselines of b+tree, alex, pgm, lipp to the others folder. Or delete the baseline related benchmark code and only run the chameleon.
• Before running the chameleon, please make sure that the computer has the necessary cuda, libtorch, and boost libraries installed.

mkdir Release
cd Release
cmake -DCMAKE_BUILD_TYPE=Release..
make

• train TSMDP index/workspace/train_TSMDP.cpp
• train DARE index/workspace/train_DARE.cpp
• run a benchmark Core components of Chameleon. Workspace is used for training Chameleon.

• Given a default parameter The constructor for the index requires one parameter. Default parameters can be used, as shown in CHA::Configuration::default_configuration.

This library provides the following classes:

• CHA::Index | index/include/Index.hpp Chameleon index structure.
• Global_Q_network | index/include/RL_network.hpp The DQN structure for DARE.
• Small_Q_network | index/include/RL_network.hpp The DQN structure for TSMDP.
• RewardScalar | index/include/RL_network.hpp A simple standardized scaling method.
• GlobalController | index/include/Controller.hpp Actor in DARE. The Implementation Process of Combining DQN and GA Algorithms.
• GlobalController::get_best_action_GA | index/include/Controller.hpp Get the best action for a given state.
• CHA::Configuration | index/include/Configuration.hpp Implementation class of parameter matrix M.
• CHA::Index::get_fanout | index/include/Index.hpp Get a fanout with the parameter matrix M with linear interpolation algorithm.
• main | index/workspace/train_DARE.hpp Train DARE with multiple threads. If there are multiple GPUs and sufficient CPU and main memory resources, the strategy for deploying processes on different GPUs can be appropriately changed.
• main | index/workspace/train_TSMDP.hpp Train TSMDP.
• experience_t | index/include/experience.hpp Implementation class of experience of DARE.
• create_dataset | include/DataSet.hpp Create dataset with local-skew distribution where the parameter skew is the skewness degree.
• local_skew | include/DataSet.hpp The function to measure the local skewness of a dataset.

[1] https://pytorch.org/