CISC $481$ Programming Assignment $3$

April $22,2024$

Assignment Objectives

In this assignment, you$$ll be implementing your own deep$1$ feed forward artificial neural network.

While completing this assignment, you$$ll learn:

$$ How to construct a directed graph representing the network

$$ How initial weight configuration affects the network$$s ability to learn

$$ A little about tuning a network$$s parameters $-$ the learning rate, which activation function

your neurons use, the number of layers in the network as well as the number of nodes in

each layer.

$$ How to implement back propagation using the L$2($loss squared$)$ function to do gradient

descent based learning.

Hexapawn

Hexapawn is a simple turn based game played on a $3\backslash$times $3$ board. Each player begins with $3$ pawns

$-$ WHITE $($MAX$)$ in the bottom row and BLACK $($MIN$)$ in the top row. Pawns can move as normal

in chess $($i$.$e$.$ white pawns can move up one square or can capture a black pawn diagonally up

one square, and black pawns can move down one square or can capture a white pawn diagonally

down one square$).$ The goal of each player is to either get one of their pawns to the other end of

the board, or to make it such that their opponent is stuck on their next move. Figure $1$ shows

the initial state of the game.

Figure $1$: The initial state of Hexapawn.

$1$It$$s a little unclear how many layers are required for an artificial neural network to qualify as a deep neural network.

You$$ll certainly be able to experiment with many layers in this project, though!

$1$