1
$\begingroup$

I have developed an RPG in likeness to the features showcased in the Final Fantasy series; multiple character classes which utilise unique action sets, sequential turn-based combat, front/back row modifiers, item usage and a crisis mechanic which bears similarity to the limit break feature.

The problem is that the greater portion of my project depends on the use of some means of Machine Learning to, in some manner, act as an actor in the game environment, however, I do not know of my options under the bare-bones environment of a command line game; I am more familiar with the use of pixel data and a neural network for action selection on a frame-by-frame basis.

Could I use reinforcement learning to learn a policy for action selection under a custom environment or should I apply a machine learning algorithm to character data, find the example outlined below, determine the best action to use on a particular turn state?

+-------+--------+--------+---------+---------+------------+---------------------+------------+--------------+--------------------+--------------+--------------+
| Level | CurrHP | CurrMP | AtkStat | MagStat | StatusList | TargetsInEnemyParty | IsInCrisis | TurnNumber   | BattleParams       |ActionOutput  | DamageOutput |
+-------+--------+--------+---------+---------+------------+---------------------+------------+--------------+--------------------+--------------+--------------+
|    65 |   6500 |    320 |      47 |      56 |          0 | Behemoth            |0           | 7            | None               |ThiefAttack   |4254          |
|    92 |   8000 |    250 |      65 |      32 |          0 | Behemoth            |1           | 4            | None               |WarriorLimit  |6984          |
+-------+--------+--------+---------+---------+------------+---------------------+------------+--------------+--------------------+--------------+--------------+

I would like to prioritise the ease of implementation of an algorithm over how optimal the potential algorithm could be, I just need a baseline to work towards. Many thanks.

$\endgroup$
  • 1
    $\begingroup$ Your third paragraph is confusing to me. Importantly, reinforcement learning (RL) is a type of machine learning, and your description of your alternative to RL "machine learning algorithm ... determine the best action" reads like it is RL too. Do you mean that you wiill personally decide what the best action is for a large number of cases, and use that as training data for a supervised learning algorithm? $\endgroup$ – Neil Slater Jun 18 at 16:19
  • $\begingroup$ @NeilSlater I understand why it's confusing, I didn't properly distinguish between actions permissible by a player in the game and actions for a learning agent as, from my knowledge, they would be indistinguishable. I am planning to both play as an expert and use collected data from a few testers as training data a supervised learning algorithm; though I do understand that this could concur bias towards particular actions. $\endgroup$ – SeunOsiko Jun 18 at 16:55
  • $\begingroup$ @SeunOsiko How good does the AI need to be? You might benefit from training the agent against itself, or rather an agent that takes the role of a potential player, with all their actions. I think in this case though, your best bet is to learn a policy for action selection. $\endgroup$ – Recessive Jun 19 at 5:07
  • $\begingroup$ @Recessive It's for an undergrad final thesis; I am more concerned about the process of developing the agent and analyzing the actions chosen on a particular game start, rather than the quality of outcome. How would I go about building an agent to learn a policy, could I use libraries such as TensorFlow or PyTorch considering the lack of visual data? $\endgroup$ – SeunOsiko Jun 19 at 12:29
  • $\begingroup$ @SeunOsiko I think you'd be best off using something like openai gym: gym.openai.com . Then, it's just a matter of how you input the data. You could actually represent this as an image if you wanted - just construct an image where each pixel is related to some action or game state. White if present or black otherwise. Keep in mind that images are ultimately arrays, so technically anything you do input can be an image if you wanted it to be. $\endgroup$ – Recessive Jun 20 at 1:55

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.