DoubleGremlin181 / RubiksCubeRL

Solving Twisty Puzzles with Reinforcement Learning

RubiksCubeRL

Solving Twisty Puzzles with Reinforcement Learning(Parallel Q-Learning)

---

Citation

@article{hukmani2021solving,
  title={Solving Twisty Puzzles Using Parallel Q-learning.},
  author={Hukmani, Kavish and Kolekar, Sucheta and Vobugari, Sreekumar},
  journal={Engineering Letters},
  volume={29},
  number={4},
  year={2021}
}

Instructions

• Install requirements
• Create a Q-Table by running pql.py on Linux or pql-windows.py on Windows.
• Download and unzip the WCA results TSV dataset
• Run scrambles_extractor.py to extract the scrambles to a csv
• Run validator.py to track the performance of your agent on official WCA scrambles
• (Optional) Solve any scramble via solver.py

Results

2x2 Pocket Rubiks Cube

Size of validation set	Average reward	Success rate	Average number of moves	Method	Training Size
10000	-86.1917	49.89	136.5806	lbl	2000000
10000	-31.9635	66.9	99.5325	none	2000000
10000	-188.8264	18.48	207.4912	ortega	2000000
10000	32.4786	85.26	53.634	lbl	3000000
10000	79.5309	99.8	21.2671	none	3000000
10000	-54.1187	53.18	127.4008	ortega	3000000
10000	77.5661	98.12	21.5351	lbl	5000000
10000	83.681	100	17.319	none	5000000
10000	31.9869	84.53	53.3884	ortega	5000000
10000	84.2557	99.77	16.512	lbl	7500000
10000	85.1919	100	15.8081	none	7500000
10000	81.5391	99.09	18.5418	ortega	7500000
10000	85.5475	100	15.4525	lbl	10000000
10000	85.5908	100	15.4092	none	10000000
10000	85.2556	99.95	15.6939	ortega	10000000

Pyraminx

Size of validation set	Average reward	Success rate	Average number of moves	Method	Training Size
10000	-221.2995	8.54	229.9249	lbl	500000
10000	-142.075	32.68	175.0818	none	500000
10000	-235.6532	4.25	239.9457	lbl	600000
10000	15.0487	80.6	66.3573	none	600000
10000	-153.2174	28.78	182.2852	lbl	700000
10000	73.6991	97.52	24.7961	none	700000
10000	-182.2324	20.06	202.493	lbl	800000
10000	82.0213	99.62	18.5949	none	800000
10000	-240.883	2.68	243.5898	lbl	900000
10000	84.6026	99.97	16.3671	none	900000
10000	-122.8171	37.63	160.8234	lbl	1000000
10000	85.615	100	15.385	none	1000000
10000	-39.7088	62.08	102.4096	lbl	1500000
10000	87.6315	100	13.3685	none	1500000
10000	51.072	88.77	38.5857	lbl	2500000
10000	89.3179	100	11.6821	none	2500000

Skewb

Size of validation set	Average reward	Success rate	Average number of moves	Method	Training Size
10000	-72.3876	54.91	127.8467	none	2000000
10000	-222.4332	8.25	230.7657	sarah	2000000
10000	81.9239	99.88	18.9549	none	3000000
10000	-125.7149	37.21	163.297	sarah	3000000
10000	86.4634	100	14.5366	none	5000000
10000	83.4465	99.18	16.7253	sarah	5000000
10000	88.1116	100	12.8884	none	7500000
10000	87.7215	99.89	13.1674	sarah	7500000
10000	88.7571	100	12.2429	none	10000000
10000	88.712	100	12.288	sarah	10000000