Design AI Systems

[Z-Puyu]

1. Explore GOAP AI mechanics.

2. Build a GOAP system and integrate with the current state machine.

3. Test the system using the current combat logic.

[Z-Puyu]

AI Decision Making Workflow?

The "Observing" Stage

Framework: Utility AI YouTube lecture.

Process

• The agent comes with a set of possible goals, for example:

  • Heal an ally;
  • Kill an opponent;
  • Protect an ally;
  • Prepare for a powerful skill.

• The agent observes the current game world and capture its environment.

• The agent starts to analyse the environment. During this process, it has some considerations which will assign an importance score to each goal based on the environment's parameters.

• Then, the goals are passed to a planner to plan out the moves.

Relevant classes

• Agent: the AI.

• Goal: a set of desired environment parameters that the agent wishes to achieve with its actions.

• Environment: keeps an internal dictionary of game world parameters.

• Parameter: a string-integer pair to describe a certain aspect of the current game world's state.

• Consideration: captures a curve to map a parameter value to its importance for each of the goals.

The "Planning" Stage

Framework: Goal-Oriented Action Planning YouTube tutorial.

Process

• The planner receives a list of goals sorted based on their importance.

• For each goal, the planner tries to search among all possible moves that can be taken to influence the current environment.

  • A move will map an environment to a different one.

  • During combat, the planner will make a move first, and predict what the opponent will react with another move.

  • Once an environment that matches the desired one of the goal, a plan will be constructed using the past moves.

  • A prudence level can be pre-defined or configured for each different agent to determine how deep its planner will search before returning all valid plans.

  • If no perfect plan can be found, those which are the closest to the desired environment will be returned as "good enough" plans.

• The planner will use the goal's importance and the "goodness" of the plan to evaluate each plan.

• All plans are then passed to the decision maker.

Relevant classes

• Planner: use BFS to search for all plans.

• Move: an IAction<Environment, Environment> to influence an environment. Each move should have a cost and an expected gain.

• Plan: contains a stack of moves. Each plan should have a cost and an expected gain.

The "Decision-Making" Stage

Process

• The decision maker receives a list of plans sorted based on their scores.

• If there is an ongoing plan currently, this plan will be re-evaluated and put back to the collection.

• One plan is picked based on the plan's score and the decision-making mode.

• The decision-making modes:

  • Logical: consider all plans which are better than half of all the choices and select one based on weighted probability.

  • Deterministic: randomly select one plan among those with the highest score.

  • Stochastic: randomly select one plan among all plans.

  • Sabotage: choose a worst plan.

• At each turn, the next move of the current plan will be executed.

Relevant classes

• DecisionMaker: a component to make the decisions.

[Z-Puyu]

Implemented the "observing" stage.

[Z-Puyu]

Implemented the "planning" and "decision-making" stages.

[Z-Puyu]

We might need someone with more experience in this to optimise the AI... :( For now, this simple framework without opponent predication should be good enough.