Page 8: Agent Intelligence: Heuristic Design
Contributed by Luiza Da Silva
Revised by Xuan Thuy Tran

[Go back to Page 7: Agent Architecture | Go on to Page 9: System Evolution]

1    Overview

In the field of Artificial Intelligence, a Heuristic is a technique of solving a problem, normally used as an aid to learning or discovery by experimental and especially trial-and-error methods. In our Acquire system, a Heuristic is going to help a player Agent make decisions about what Actions to take. Heuristics are necessary in Acquire because Agents have imperfect knowledge about the game, because a given Agent does not have full knowledge of other Agents' plans, number and type of stocks, amount of money, etc. Examples of situations in the game that contribute to the condition of imperfect knowledge are:

• Tiles are picked at random in the beginning of the game and at the end of each turn;
• the Agent does not keep track of what the other player Agents possess (stock, money, or tiles);
• there is a large amount of Action possibilities that Agents can choose from at any given point in the game.

The huge number of possible moves the agent can choose from is the reason why it needs the structure described in the Agent Architecture section. An Agent has to devise a way to select one move among all the others, based on a fixed goal that it plans to reach at the end of the game. The selection of an Action involves comparing states of the world, and analyzing which state maximizes an Agent's Heuristic's goals.

The Heuristic is responsible for considering all possible moves the Agent can currently make (i.e., all the tiles it can currently place and all stocks it can currently buy) and determine which Action has the most benefit. The Heuristic has access to the current world State as well as the current Agent state. In the current general Heuristic which all Agents within the system use, the guiding strategy is to take the action which increases the overall or expected assets of the current Agent. The so-called "deep goal" of each Agent is to have the most possible stock in the largest chain when the end of the game occurs. This takes place in two phases:

• Placing a tile: Comparison between states is based on mathematical function manipulation of several key parameters that are important to the domain of Acquire. These are new-chain, merge-chains, grow-chain, and solitary-tile. They are described in further detail below in part 4.
  
  
• Buying stock: To fulfill the Agent's deep goal, the Heuristic considers all chains currently on the Board and estimates each one's maximum expected size at the end of the game, buying stock in that one. In the case of a tie, the Heuristic recommends buying in a chain the Agent already owns.

2    General Heuristic Architecture

The General Heuristic consists of several components:

• Heuristic - takes the Board State and the Agent State, calculates the best choice move for the Agent, and returns this result.
• Parameter Set - a list of all relevant parameters about the world and the player's position in the game; these can be set with actual values or with the results of a function from the Function Set; see section 4 of this document below for more details.
• Function Set - the mathematical functions (if any) which a certain Heuristic may need to determine the weights for some parameters in the Parameter Set.

3    Supporting Classes

• State - a class which encapsulates the state of the world as it currently is, the most useful information, to use as a container to pass the state between objects, like in this case, between Agents and Heuristics. The State class can be found in the Acquire Engine Architecture section.

4    Parameter Set

The parameters used by the Heuristic to determine overall benefit of each possible Action to the Agent player are described below. They represent the four possible effects of any given "Place-tile" action; their order basically determines the strategy that Agent will use. In the General Heuristic, the order is "merge-chains, grow-chain, new-chain, and solitary-tile". In the "Buy-stock" portion of evaluation, the Heuristic attempts to buy stock in the chain which it can calculate to have the largest expected size at the end of the game.