Skip to content

List of Agents

[source code]

AgentHuman

connectFourLab.game.agents.basicAgents.AgentHuman()

Human Agent represents the user as a player

When a UI implements RunGame it will assign a function to get_input which will be called each turn. The function should return the column choosed by the user.

[source code]

AgentRandom

connectFourLab.game.agents.basicAgents.AgentRandom()

Random agent, play a valid column chosen randonly

[source code]

AgentNegamax

connectFourLab.game.agents.negamax.AgentNegamax()

Negamax Agent

This agent uses the negamax algorithm which is a simplified version of the minimax algorithm.

Negamax is based on the observation that max(a,b) = -min(-a,-b)

This agent uses negamax in all actions using the maximum depth of 5.

It also uses a zobrist hash table to store all searches in order to be more efficient.

[source code]

AgentSimulation

connectFourLab.game.agents.basicAgents.AgentSimulation()

Simulation agent

When taking an action this agent simulates a number of games for each available column and then choose the best probabilistic option.

The best probabilistic option is the best score of a column, in a 100 simulated games.

Each simulation consists in simulate every turn randomly for both players until the simulation ends in a terminal state. Then it will return 1 for victory, -1 for defeat or 0 in case of a draw.

[source code]

AgentSimulationTL

connectFourLab.game.agents.basicAgents.AgentSimulationTL()

Simulation strategy with time management

This agent act by simulating a number of games for each available column and then choosing the best probabilistic option.

Similar with the AgentSimulation. Differentiating itself by instead of simulating a fixed number of games, this agent manage the time available through the game, simulating the maximum amount of games (equaly distributed to all possible columns) without letting the time run out and thus being able to handle limited time games.

In unlimited time games the agents takes the maximum of 20 seconds per turn before returning the choice.

[source code]

AgentMonteCarlo

connectFourLab.game.agents.basicAgents.AgentMonteCarlo()

Monte Carlo agent.

This agent applies the Monte Carlo Tree Search method. The method consists in search the possibilities of the board evaluating each stage of the board (rollout), but different from a minimax tree search this search don't go into all the possibilities, it uses the UCB1 algorithm to measure how much of the search effort goes into exploiting promissing branches and exploring less prommissing branches.

This agent evaluate the rollout with a simulation of 100 games per board state.

[source code]

AgentMCTSNN

connectFourLab.game.agents.mctsnn.AgentMCTSNN(model_file=None, model=None)

Agent Monte Carlo Tree Search with Neural Network evaluation.

This agent inherit from AgentMonteCarlo using the same process but evaluating the rollouts by using a trained neural network model.

This agent uses the Node NodeMCTSNN which evaluate the rollout score by predicting the board state in a trained model.

Arguments

  • model_file - str; path to the '.h5' model file;
  • model - str, loaded keras.models.Model object;

Exceptions

  • MissingModel - raise when creating a new instance, if both model_file and model are None.