c4a0: Connect Four Alpha-Zero

An alpha-zero-style Connect Four neural network trained via self play.

• c4.py: Connect four game logic
• nn.py: Neural net architecture
• mcts.py: Monte carlo tree search implementation
• self_play.py: Multiprocessing self play training sample generation
• training_py: Generational training implementation
• tournament.py: Round-robin multi-model tournament implementation

Usage

poetry install --no-root to install deps.

poetry run src/main.py train to train:

• Stores training state in training/ folder

poetry run src/main.py tournament --gen-id 22 --gen-id 21 --gen-id 18:

• Runs a tournament with the given generations

Results

After training for ~12 hours with an RTX 3090, we achieved the following results:

• Number of generations trained: 22
• Total training positions: 2,361,492

Final ELO table

	Player	ELO	Rank
0	gen20	1811.67	1
1	gen4	1594.95	2
2	gen15	1563.46	3
3	gen11	1557.35	4
4	gen22	1546.2	5
5	gen12	1543.55	6
6	gen14	1543.14	7
7	gen13	1541.27	8
8	gen16	1532.4	9
9	gen17	1516.62	10
10	gen10	1513.8	11
11	gen18	1510.89	12
12	gen2	1503.38	13
13	gen19	1500.64	14
14	gen8	1496.41	15
15	gen7	1487.36	16
16	gen6	1464.17	17
17	gen5	1443.68	18
18	gen3	1417.44	19
19	gen21	1393.73	20
20	gen1	1385.84	21
21	uniform	1385.49	22
22	gen9	1378.7	23
23	random	1367.88	24

Training gen 1 from 0. Tournament results:

Player	Score
gen1	46
uniform	41
random	33

Training gen 2 from 1. Tournament results:

Player	Score
gen2	107.5
gen1	45.5
uniform	45
random	42

Training gen 3 from 2. Tournament results:

Player	Score
gen2	129
gen3	113
gen1	70.5
random	52.5
uniform	35

Training gen 4 from 2. Tournament results:

Player	Score
gen4	135
gen2	94
gen3	92.5
gen1	48
random	30.5

Training gen 5 from 4. Tournament results:

Player	Score
gen2	100
gen4	100
gen3	100
gen5	70
gen1	30

Training gen 6 from 2. Tournament results:

Player	Score
gen3	100
gen6	100
gen4	80
gen2	80
gen5	40

Training gen 7 from 3. Tournament results:

Player	Score
gen4	120
gen6	100
gen7	80
gen3	60
gen2	40

Training gen 8 from 4. Tournament results:

Player	Score
gen4	120
gen8	90
gen6	80
gen7	70
gen3	40

Training gen 9 from 4. Tournament results:

Player	Score
gen4	120
gen8	90
gen9	80
gen6	60
gen7	50

Training gen 10 from 4. Tournament results:

Player	Score
gen8	100
gen9	100
gen10	80
gen4	80
gen6	40

Training gen 11 from 8. Tournament results:

Player	Score
gen11	130
gen8	80
gen9	70
gen10	60
gen4	60

Training gen 12 from 11. Tournament results:

Player	Score
gen12	120
gen11	110
gen8	80
gen9	50
gen10	40

Training gen 13 from 12. Tournament results:

Player	Score
gen12	130
gen13	120
gen11	70
gen8	70
gen9	10

Training gen 14 from 12. Tournament results:

Player	Score
gen12	110
gen13	100
gen14	80
gen11	80
gen8	30

Training gen 15 from 12. Tournament results:

Player	Score
gen15	110
gen11	80
gen12	80
gen13	80
gen14	50

Training gen 16 from 15. Tournament results:

Player	Score
gen15	100
gen12	80
gen13	80
gen16	80
gen11	60

Training gen 17 from 15. Tournament results:

Player	Score
gen15	110
gen12	100
gen16	70
gen17	70
gen13	50

Training gen 18 from 15. Tournament results:

Player	Score
gen12	100
gen15	100
gen16	70
gen17	70
gen18	60

Training gen 19 from 12. Tournament results:

Player	Score
gen12	90
gen15	90
gen17	80
gen19	80
gen16	60

Training gen 20 from 12. Tournament results:

Player	Score
gen20	120
gen12	80
gen17	70
gen15	70
gen19	60

Training gen 21 from 20. Tournament results:

Player	Score
gen20	130
gen15	90
gen12	80
gen21	50
gen17	50

Training gen 22 from 20. Tournament results:

Player	Score
gen20	150
gen22	80
gen15	80
gen12	60
gen21	30