In this project, I developed an adversarial search agent to play the game "Isolation". Isolation is a deterministic, two-player game of perfect information in which the players alternate turns moving a single piece from one cell to another on a board. Whenever either player occupies a cell, that cell becomes blocked for the remainder of the game. The first player with no remaining legal moves loses, and the opponent is declared the winner. These rules are implemented in the isolation.Board class provided in the repository.
This project uses a version of Isolation where each agent is restricted to L-shaped movements (like a knight in chess) on a rectangular grid (like a chess or checkerboard). The agents can move to any open cell on the board that is 2-rows and 1-column or 2-columns and 1-row away from their current position on the board. Movements are blocked at the edges of the board (the board does not wrap around), however, the player can "jump" blocked or occupied spaces (just like a knight in chess).
Additionally, agents will have a fixed time limit each turn to search for the best move and respond. If the time limit expires during a player's turn, that player forfeits the match, and the opponent wins.
I implemented the following functions:
-
MinimaxPlayer.minimax(): implement minimax search -
AlphaBetaPlayer.alphabeta(): implement minimax search with alpha-beta pruning -
AlphaBetaPlayer.get_move(): implement iterative deepening search -
custom_score(): implement your own best position evaluation heuristic -
custom_score_2(): implement your own alternate position evaluation heuristic -
custom_score_3(): implement your own alternate position evaluation heuristic
You may write or modify code within each file (but you must maintain compatibility with the function signatures provided). You may add other classes, functions, etc., as needed, but it is not required.
The Project Assistant sandbox for this project places some restrictions on the modules available and blocks calls to some of the standard library functions. In general, standard library functions that introspect code running in the sandbox are blocked, and the PA only allows the following modules random, numpy, scipy, sklearn, itertools, math, heapq, collections, array, copy, and operator. (Modules within these packages are also allowed, e.g., numpy.random.)
The following example creates a game and illustrates the basic API. You can run this example by activating your aind anaconda environment and executing the command python sample_players.py
from isolation import Board
from sample_players import RandomPlayer
from sample_players import GreedyPlayer
# create an isolation board (by default 7x7)
player1 = RandomPlayer()
player2 = GreedyPlayer()
game = Board(player1, player2)
# place player 1 on the board at row 2, column 3, then place player 2 on
# the board at row 0, column 5; display the resulting board state. Note
# that the .apply_move() method changes the calling object in-place.
game.apply_move((2, 3))
game.apply_move((0, 5))
print(game.to_string())
# players take turns moving on the board, so player1 should be next to move
assert(player1 == game.active_player)
# get a list of the legal moves available to the active player
print(game.get_legal_moves())
# get a successor of the current state by making a copy of the board and
# applying a move. Notice that this does NOT change the calling object
# (unlike .apply_move()).
new_game = game.forecast_move((1, 1))
assert(new_game.to_string() != game.to_string())
print("\nOld state:\n{}".format(game.to_string()))
print("\nNew state:\n{}".format(new_game.to_string()))
# play the remainder of the game automatically -- outcome can be "illegal
# move", "timeout", or "forfeit"
winner, history, outcome = game.play()
print("\nWinner: {}\nOutcome: {}".format(winner, outcome))
print(game.to_string())
print("Move history:\n{!s}".format(history))