How to Use DeepCubeA for Problem Solving

DeepCubeA is a powerful framework designed for solving various combinatorial puzzles like the Rubik’s Cube, 15-puzzle, Lights Out, and more using Deep Reinforcement Learning and A* search algorithms. In this guide, we’ll walk you through the setup process, training, and enhancing the performance of DeepCubeA to effectively tackle these puzzles.

Setup Instructions

• Ensure you have Python 3.7.2 installed on your machine.
• Check the requirements.txt for the necessary Python packages.
• Install these packages using pip or conda.
• Navigate to the DeepCubeA directory and execute the following command to set up the environment:

  source setup.sh

Training and A* Search

To train DeepCubeA and implement A* search, follow these steps:

Training Commands

Use the following command format to train your model for the 15-puzzle:

python ctg_approxavi.py --env puzzle15 --states_per_update 50000000 --batch_size 10000 --nnet_name puzzle15 --max_itrs 1000000 --loss_thresh 0.1 --back_max 500 --num_update_procs 30

Solving with A* Search

After the training process, utilize A* search to solve the puzzle:

python search_methodsastar.py --states datapuzzle15testdata_0.pkl --model saved_modelspuzzle15current --env puzzle15 --weight 0.8 --batch_size 20000 --results_dir resultspuzzle15 --language cpp --nnet_batch_size 10000

Comparing Solutions

If you want to evaluate the effectiveness of your solution, compare it to the shortest path:

python scriptscompare_solutions.py --soln1 datapuzzle15testdata_0.pkl --soln2 resultspuzzle15results.pkl

Improving Results

To enhance your model’s performance during training and search, consider the following adjustments:

• Increase --batch_size and --states_per_update for better results during approximate value iteration (AVI).
• Lower the --loss_thresh to trigger updates more frequently.
• Add states encountered through greedy best-first search (GBFS) to the training set.
• Adjust the weight on path cost during A* search to improve outcomes.

Creating New Problem Solvers

DeepCubeA can be adapted to solve new puzzles by implementing your own environment. Follow these steps:

• Implement the abstract methods found in environments/environment_abstract.py.
• Edit utils/env_utils.py to register your environment.
• Test the implementation using tests/timing_test.py to ensure everything functions properly.

Parallelism in Training

To speed up the training and solving processes, leverage multi-CPU and multi-GPU setups:

• Set the number of workers for approximate value iteration using --num_update_procs.
• Control the number of GPUs by setting the CUDA_VISIBLE_DEVICES environment variable, for example:

  export CUDA_VISIBLE_DEVICES=0,1,2,3

Memory Management

While acquiring training data and solving with A* search, you may need to reduce the batch size if memory issues arise. Adjust the following settings:

• --update_nnet_batch_size for avi.py
• --nnet_batch_size for astar.py

Compiling C++ for A* Search

If you’re interested in utilizing C++ for A* search, compile the necessary files by navigating to the cpp directory and running:

cd cpp

make

If you face challenges with the C++ version, you can switch to the Python implementation by modifying the --language parameter in search_methods/astar.py from --language cpp to --language python.

Troubleshooting

Should you encounter issues while using DeepCubeA, consider these troubleshooting steps:

• Verify your Python environment and installed package versions match the requirements.
• Ensure all necessary files are in the correct directories, especially when executing commands.
• Reduce batch sizes or update steps if you experience memory-related errors.
• If you have further questions or need assistance, reach out to Forest Agostinelli at foresta@cse.sc.edu.

For more insights, updates, or to collaborate on AI development projects, stay connected with fxis.ai.

Conclusion

At fxis.ai, we believe that such advancements are crucial for the future of AI, as they enable more comprehensive and effective solutions. Our team is continually exploring new methodologies to push the envelope in artificial intelligence, ensuring that our clients benefit from the latest technological innovations.