From ee216c008a188a9436fedb85c70ee5d1719733b1 Mon Sep 17 00:00:00 2001 From: tslil clingman Date: Sun, 15 Jan 2023 16:03:37 +0100 Subject: new neural network arch (faster + better) & minor changes + fixes Gone is the convolutional neural network, for it turns out not only is it more difficult to train, but all of the extra information about board layers didn't make much of a difference at this size. So cnn1986 has been replaced by nn1986, a standard, two-layer, dense nn configured as a binary classifier and (mis)used in that capacity. Note: total number of parameters is unchanged. HARK: this new nn exposes a bug somewhere in ctak. Run ctlm with self-play to see the completely borked board state at the end. --- README.md | 24 ++++++++++++++++++------ 1 file changed, 18 insertions(+), 6 deletions(-) (limited to 'README.md') diff --git a/README.md b/README.md index ea0e172..9f18cad 100644 --- a/README.md +++ b/README.md @@ -12,7 +12,7 @@ Details forthcoming, but at a glance: ## Building for the native platform - make native + make native ## Building for the Raspberry Pi Zero @@ -20,17 +20,29 @@ Download and extract a recent version of buildroot into the working directory. Edit `BUILDROOT_DIR=buildroot-2020.11.1` in `Makefile` to point to the extracted directory. - make pi + make pi ## The computer opponent -### Adversarial tree search implementation +Standard adversarial tree search (α-β negamax) with iterative deepening, using a neural network evaluation function for leaves and also transposition tables implemented using Zobrist hasing and a chaining hash table. -Details coming soon +The architecture of the neural network is a standard, two-layer dense classification network +``` +_________________________________________________________________ + Layer (type) Output Shape Param # +================================================================= + input_1 (InputLayer) [(None, 28)] 0 -### The convolutional neural network cnn1986 + dense (Dense) (None, 64) 1856 -Details coming soon + dense_1 (Dense) (None, 2) 130 + +================================================================= +Total params: 1,986 +Trainable params: 1,986 +Non-trainable params: 0 +``` +which was trained on the binary classification problem of predicting the winner from a given board state. As input the network is fed the top layer of the board only (see nn1986.c for details) as well as `flats used/flats remaining` fractions for both players and a single float indicating the parity of the board. At the time of training, on the dataset given by `resources/extract.sh`, this achieves ~82% accuracy on the validation set. ## License -- cgit v1.2.3