Message Boards Message Boards

First forays into game design and agent reasoning about uncertainty

Posted 9 years ago

Over the past year, I've found it increasingly more difficult to find video games that can keep me entertained. This has forced me to focus more on making my own. Making games was the main reason I got into programming as a teenager (and a popular reason for many, I believe), so it's nice to return to it. For my first game last year, I wanted to make something interesting as quickly as possible. This led me to make a small puzzle game. I prototyped it in Mathematica, and now you can play the JavaScript version here ( www.gatheredgame.com ).

That was a fun exercise, but moving forward, I wanted to create less abstract systems. I'm interested in small fictional worlds, emphasizing intricacy and responsiveness. This would enable player-driven stories, without having to hand-code every possible outcome and without limiting myself to the building and strategy genres. Perhaps things like Dwarf Fortress, but more accessible and varied.

I decided a key to accomplishing this was to more closely tie together the creative writing and programming parts of my design process. I knew that the final result wouldn't hold my attention long enough to be worth the effort if I took the industry approach to game design by starting from an existing genre or game and changing the setting or adding a new feature. Some independent games have been successful by starting from a novel interaction mechanic, and then exploring the language of puzzles that can be generated from that grammar. A prominent example is Portal. I think starting from random creative writing exercises is more fun, but in my initial attempts I had to throw away many details to reduce the results down to something I could start coding. That made the design process feel very inefficient. I want coding a game to feel like creative writing, and I want the output to be interesting from the first few lines until the final result.

Ever since reading NKS, I've been struck by how interesting a random cellular automata browser can be given how little code is required to make one. For example, try enabling rules below until interesting behavior is produced (click the squares to the right of the arrows).

SeedRandom[1];
init = RandomInteger[1, 100]; rule = 
 Table[False, {2^5}]; Row@{Column@
   Table[Row@{Row[
       IntegerDigits[a - 1, 2, 
         5] /. {0 -> 
          Graphics[{White, EdgeForm[{Thin, Black}], Rectangle[]}, 
           ImageSize -> 10], 
         1 -> Graphics[{Black, EdgeForm[{Thin, Black}], Rectangle[]}, 
           ImageSize -> 10]}], "\[Rule]", 
      With[{a = a}, 
       Dynamic@Graphics[
         Button[If[
           rule[[a]], {{Black, EdgeForm[{Thin, Black}], 
             Rectangle[]}}, {{White, EdgeForm[{Thin, Black}], 
             Rectangle[]}}], rule[[a]] = ! rule[[a]]], 
         ImageSize -> 10]]}, {a, 2^5}], Spacer@10, 
  Dynamic@ArrayPlot[
    CellularAutomaton[{FromDigits[Reverse@Boole@rule, 2], 2, 2}, init,
      150], ImageSize -> 350]}

enter image description here

So I decided the easiest way to try and take advantage of the fundamental phenomenon that complex behavior isn't difficult to produce, while simultaneously focusing on concrete objects, was to experiment with association automata. With "association" referring to Wolfram's key/value dictionaries.

I like starting from random story ideas. I currently pull random story tropes from the TV Tropes idea generator to get started. Sometimes I mix in a few random Wikipedia articles.

randomStoryIdea[] := 
 DynamicModule[{elements = 
    Import["http://tvtropes.org/pmwiki/storygen.php", "Data"][[2, 2, 
     2]]}, Dynamic@
   Grid[MapIndexed[{#[[1]], 
       Hyperlink[#[[2]], 
        "http://tvtropes.org/pmwiki/pmwiki.php/Main/" <> 
         StringReplace[#[[2]], Except@LetterCharacter -> ""]], 
       Button["New", 
        elements[[#2[[1]]]] = 
         Import["http://tvtropes.org/pmwiki/storygen.php", "Data"][[2,
           2, 2, #2[[1]]]]]} &, elements], Alignment -> Left]]

Here is a random seed I'm working with now:

enter image description here

Then I start to stitch some of the elements together in my head. An evil poacher at a lighthouse made me think of a fisherman, perhaps a criminal hiding in a lighthouse who is fishing to survive. Chain of harm and monster adventurers made me think that perhaps his fishing activities are damaging a community of marine life under the dock, and some crabs attempt to stop him. Now I'll start defining some entities and properties. I start with the properties that I think someone would consider most interesting at the end of a game. Health and location of main characters, how many catches did the fisherman complete, what is the state of the crab town?

entities = <|
  "fisherman" -> <|"health" -> "healthy", "location" -> "dock", 
    "catches" -> 0|>,
  "hero crab" -> <|"health" -> "healthy"|>,
  "assistant crab" -> <|"health" -> "healthy"|>,
  "crab town" -> <|"population" -> 50, "crime rate" -> .05|>
  |>

Then I define some rules to update those properties, referring to new properties that will influence them. In this way I essentially define the game world by starting from final effects and working backward by defining causes. The entities and properties are defined as an association of associations. The rules are defined as a list of lists that is then converted to associations with some code. Each sublist represents one rule that updates one property of one entity. The first two items of the sublist are the entity and property being updated, and the last item is a pure function that defines the behavior of the rule. Three values are passed to the rule when it is used. Slot[1] is the current value of the property being updated. Slot[2] is the current entity being updated, so I can refer to other properties of that entity easily with #2["property"]. Slot[3] is the association containing the current state of all entities, so I can refer to properties of the fishing rod within a rule that updates the fisherman's catch count with #3["fishing rod","fish on line"].

rules = {
  {"fisherman", "health", Which[
     #2@"pacemaker" === "off", "dead",
     #2@"fear level" >= 30, "dead",
     #2@"location" === "ocean", "dead",
     #2@"blood level" < 30 && # === "healthy", "injured",
     #2@"blood level" <= 0, "dead",
     #2@"hunger level" >= 50, "dead",
     # === "injured" && #2@"time without bleeding" >= 20, "healthy",
     True, #] &},
  {"crab town", 
   "population", # + If[RandomReal[] < #2@"crime rate", -1, 0] + 
     If[RandomReal[] < #/4*.005, 1, 0] &}
  }

Before we look at the behavior of the system, we need to define initial values for any new properties we reference in the rules. Here is some code to find those.

extractVars[fun_, scopedEntity_] := 
 Join[Cases[fun, #2[prop_] :> {scopedEntity, prop}, \[Infinity]], 
  Cases[fun, #3[globalEntity_, prop_] :> {globalEntity, 
     prop}, \[Infinity]], 
  Cases[fun, 
   prop[globalEntity_String, prop_String] :> {globalEntity, 
     prop}, \[Infinity]]]

Complement[
 Union[Join @@ (Prepend[extractVars[#3, #], {#, #2}] & @@@ rules)], 
 Join @@ KeyValueMap[Thread[{#, Keys@#2}] &, entities]]

The following code expands the rules into associations and allows for simple one-level, OO-style inheritance of properties and rules by using an "is a" property on an entity. This example doesn't utilize that, but the code to apply rules assumes they have been transformed into associations.

(*currently only supports one level of inheritance. inherited rules \
lose ordering*)
fullEntities = entities;
Select[fullEntities, KeyExistsQ@"is a"] // 
   KeyValueMap[{#, #2@"is a"} &] // 
  Apply[{entity, parent} \[Function] 
     AssociateTo[fullEntities, 
      entity -> 
       Join[fullEntities@parent, fullEntities@entity]], #, {1}] &;

fullRules = 
  Select[entities, KeyExistsQ@"is a"] // 
    KeyValueMap[Cases[rules, {#2@"is a", a__} :> {#, a}] &] // 
   Select[<|"entity" -> #, "property" -> #2, "rule" -> #3, 
        If[Length@{##} > 3, "probabilities" -> {##}[[4]], 
         "probabilities" -> Null]|>, # =!= Null &] & @@@ 
     Join[Catenate@#, rules] &;

Then to step through the evolution of the system, we can use code like the following. In my example the only changes come from a gradually changing population due to birth and crime, so you can use the input field above the step button to take more than one step at a time:

applyRules[entities_, rules_] := 
 Fold[ReplacePart[#, {#2@"entity", #2@"property"} -> #2[
       "rule"][#[#2@"entity", #2@"property"], #[#2@"entity"], #]] &, 
  entities, rules]

play[] := (state = fullEntities; steps = 0; stepSize = 1; 
  Grid[{{Column@{InputField[Dynamic@stepSize, Number, FieldSize -> 4],
        Button["Step", 
        state = Nest[applyRules[#, fullRules] &, state, stepSize]; 
        steps += stepSize], Dynamic@steps}, 
     Dynamic@Column@Normal[state]}}, Frame -> All, 
   Alignment -> {Center, Center}])

play[]

To help decide what rules to add next, I have some code that visualizes which properties influence other properties as a graph, and some code that sorts properties by how many other properties they influence. This makes it easy to add rules that have the greatest downstream effects on the system.

Graph[Flatten[
  Thread[Union@extractVars[#[[3]], #[[1]]] \[DirectedEdge] #[[;; 2]], 
     List, 1] & /@ rules, 1], VertexLabels -> "Name", 
 PlotRangePadding -> 1]

Module[{counts = 
   Counts[Join @@ (Union@extractVars[#3, #] & @@@ fullRules)], 
  counts2 = Counts[Join @@ (extractVars[#3, #] & @@@ fullRules)]}, 
 Complement[
    Join[Join @@ KeyValueMap[Thread[{#, Keys@#2}] &, fullEntities], 
     Keys@counts], Values@fullRules[[All, {"entity", "property"}]]] //
    Map[{#, Lookup[counts, Key@#, 0], Lookup[counts2, Key@#, 0]} &] //
   SortBy[-Last@# &]]

enter image description here

As you start expanding your system, you'll very quickly realize that defining rules for the behavior of "intelligent" entities in your system is very tedious. I've started investigating some ways to simplify this. We'll use the fisherman as an example, but I plan to use the same techniques for the crabs and others. We can just treat his choice of what action to perform as a search and optimization problem. By giving him a utility or objective property to maximize like "happiness", we can just have him search through a few steps of simulating the system for various possible actions and choose the one that leads to the most happiness.

optimize[property_, choices_, objective_, entities_, rules_, 
  searchDepth_] := 
 Module[{rulesSafe = 
    Select[rules, property =!= prop[#entity, #property] &]}, 
  First@MaximalBy[
    Table[{c, 
      Module[{newState = 
         applyRules[ReplacePart[entities, List @@ property -> c], 
          rulesSafe]}, 
       If[searchDepth > 1, 
        Last@optimize[property, choices, objective, newState, 
          rulesSafe, searchDepth - 1], newState @@ objective]]}, {c, 
      entities @@ choices}], Last]]

You would use that function similar to the following:

{"fisherman", "action", 
 First@optimize[prop["fisherman", "action"], 
    prop["fisherman", "possible actions"], 
    prop["fisherman", "happiness"], #3, fullRules, 5] &}

I gave the fisherman a few possible actions (sitting, standing, casting, fishing, reeling, unhooking) and defined which ones are accessible under which conditions with nested Which expressions. Then I defined his happiness to increase by one whenever he unhooks a fish.

Now we immediately run into a problem where he has no way to prioritize unhooking a fish now when he could just wait a turn and unhook it then. If he is recursively searching two moves ahead, then both unhooking now then waiting or waiting now then unhooking both return the same happiness. So depending on search order, he might just choose to wait every single step. One simple fix for this example is to just multiply his happiness by something like 1.01 at each step. Then he will prioritize unhooking the fish as soon as possible, because ((x+1)1.01)1.01 is greater than ((1.01x)+1)1.01. The sooner he gets happiness increases, the greater the effect of the ongoing multiplier. You could also have his happiness be diminished based on search depth in the optimization recursion, have the formula directly incorporate the amount of time that has passed in the game, etc.

entities = <|
   "fisherman" -> <|"health" -> "healthy", "location" -> "dock", 
     "catches" -> 0, "action" -> "sitting", "happiness" -> 50, 
     "possible actions" -> {"sitting", "casting", "standing"}, 
     "blood level" -> 100, "fear level" -> 0, "hunger level" -> 0, 
     "pacemaker" -> "on", "time without bleeding" -> 0, 
     "catching fish" -> False|>,
   "hero crab" -> <|"health" -> "healthy"|>,
   "assistant crab" -> <|"health" -> "healthy"|>,
   "crab town" -> <|"population" -> 50, "crime rate" -> .05|>,
   "fishing rod" -> <|"crab on line" -> False|>,
   "dock" -> <|"trip wire" -> "not set"|>
   |>;

rules = {
   {"fisherman", "action", 
    First@optimize[prop["fisherman", "action"], 
       prop["fisherman", "possible actions"], 
       prop["fisherman", "happiness"], #3, fullRules, 5] &},
   {"fisherman", 
    "catching fish", #2@"action" === "unhooking" && #3["fishing rod", 
       "crab on line"] &},
   {"crab town", "crime rate", .05 + .05*#3["fisherman", "catches"] &},
   {"fisherman", "location", Which[
      #2@"fear level" >= 20 && # === 
        "dock" && #3["dock", "trip wire"] === "set", "ocean",
      #2@"action" === "going to lighthouse", "lighthouse",
      #2@"action" === "going to dock", "dock",
      True, #] &},
   {"fisherman", "possible actions", Which[
      #2@"location" === "dock", Which[
       #2@"action" === "sitting", {"sitting", "casting", "standing"},
       #2@"action" === "casting", {"fishing"},
       #2@"action" === "fishing", {"fishing", "reeling"},
       #2@"action" === "reeling", {"sitting", "unhooking"},
       #2@"action" === "unhooking", {"sitting"},
       #2@"action" === "standing", {"sitting", "going to lighthouse"},
       True, #],
      True, #] &},
   {"fisherman", "health", Which[
      #2@"pacemaker" === "off", "dead",
      #2@"fear level" >= 30, "dead",
      #2@"location" === "ocean", "dead",
      #2@"blood level" < 30 && # === "healthy", "injured",
      #2@"blood level" <= 0, "dead",
      #2@"hunger level" >= 50, "dead",
      # === "injured" && #2@"time without bleeding" >= 20, "healthy",
      True, #] &},
   {"fisherman", "happiness", 
    1.01 (# + If[#2@"catching fish", 1, 0] - 
        If[#2@"health" === "injured", 20, 0]) &},
   {"fisherman", "catches", If[#2@"catching fish", # + 1, #] &},
   {"fishing rod", "crab on line", Which[
      #3["fisherman", "action"] === "fishing" && RandomReal[] < .1, 
      True,
      #3["fisherman", "action"] === "unhooking", False,
      True, #] &},
   {"crab town", 
    "population", # + If[RandomReal[] < #2@"crime rate", -1, 0] + 
      If[RandomReal[] < #/4*.005, 1, 0] &}
   };

enter image description here

So now if he was guaranteed to catch a fish within the search depth of his optimization (I'm using 5 for now), he would behave as desired. The problem is that for each step that he is fishing, he only has a 10% chance of a fish biting. This means that almost 9 of 10 times, his optimization search will return no better options than doing nothing. He needs to understand that the chance of catching a fish is always better than doing nothing. I'm experimenting with just manually annotating any rules that depend on randomness. Then I define a new optimization search function that can read these annotations, split the search into multiple branches when necessary, and return the expected happiness from the split. In our example, when he is fishing and no fish has bitten yet, it will return a happiness of .1(the happiness from continuing the search assuming a fish bites)+.9(the happiness from continuing the search assuming a fish doesn't bite).

{"fishing rod", "crab on line", 
 Which[#3["fisherman", "action"] === "fishing" && RandomReal[] < .1, 
   True, #3["fisherman", "action"] === "unhooking", False, 
   True, #] &, {{#3["fisherman", "action"] === 
      "fishing" && ! # &, {{.1, True}, {.9, False}}}}}

relevantProbs[rule_, state_] := 
 If[KeyExistsQ[rule, "probabilities"], 
  Select[rule[
    "probabilities"], #[[1]][state[rule@"entity", rule@"property"], 
     state[rule@"entity"], state] &], {}]

applyRulesProb[entities_, rules_] := 
 Fold[Function[{states, rule}, 
   Function[state, 
     If[Length@relevantProbs[rule, state[[2]]] > 0, 
      relevantProbs[rule, state[[2]]][[1, 2]] // 
        Map[Function[
          branch, {branch[[1]]*state[[1]], 
           ReplacePart[
            state[[2]], {rule@"entity", rule@"property"} -> 
             branch[[2]]]}]] // Sequence @@ # &, {state[[1]], 
       ReplacePart[
        state[[2]], {rule@"entity", rule@"property"} -> 
         rule["rule"][state[[2]][rule@"entity", rule@"property"], 
          state[[2]][rule@"entity"], state[[2]]]]}]] /@ states], {{1.,
     entities}}, rules]

optimizeProb[property_, choices_, objective_, entities_, rules_, 
  searchDepth_] := 
 Module[{rulesSafe = 
    Select[rules, property =!= prop[#entity, #property] &]}, 
  First@MaximalBy[
    Table[{c, 
      Module[{newState = 
         applyRulesProb[ReplacePart[entities, List @@ property -> c], 
          rulesSafe]}, 
       If[searchDepth > 1, 
        Function[branch, 
           branch[[1]]*
            Last@optimizeProb[property, choices, objective, 
              branch[[2]], rulesSafe, searchDepth - 1]] /@ newState //
          Total, Function[branch, 
           branch[[1]]*branch[[2]] @@ objective] /@ newState // 
         Total]]}, {c, entities @@ choices}], Last]]

He now displays the desired behavior of choosing to start fishing immediately, waiting until a fish bites, then reeling and unhooking, then continuing.

enter image description here

Right now, the fisherman has perfect knowledge of the system. This is because I pass in the full rules and full set of entities of the system to the optimization search. An interesting next step as I add rules and properties to the system will be to use a subset of the entities and rules that represent his limited mind and knowledge. Then I can add rules that update his knowledge and even his happiness function over time to make him learn. For example, say whether or not a fish bites is actually deterministic based on more detailed rules for crab behavior. However, let's says the fisherman's copy of the system rules excludes those and instead just assumes a 30% random chance of catching a fish. Further, assume the 30% expectation is represented as a property of the fisherman (say "fishing optimism"). Then we can easily add a rule that increases his optimism whenever he catches a fish, and decreases it whenever he doesn't. Then say we add time of day to the system, and expand his happiness function to decrease when he stays out too late fishing. Then he will automatically choose to stay out a little later when he has been a catching a lot of fish compared to when he hasn't been catching many.

That's just one possible extension I've considered. I don't know if this approach will ever lead to me making a commercially successful game, but it's holding my attention for now.

POSTED BY: Michael Hale
5 Replies

enter image description here - you earned "Featured Contributor" badge, congratulations !

Dear @Michael Hale, this is a great post and it has been selected for the curated Staff Picks group. Many people will receive a monthly digest featuring you post. Your profile is now distinguished by a "Featured Contributor" badge and displayed on the "Featured Contributor" board.

POSTED BY: EDITORIAL BOARD

Great post, @Michael Hale ! - And a great discussion. I've added a few animated GIFs. I am sure @Rob Lockhart, who is in the game industry, would be curious to take a look. And you guys would probably be curios to know about his Important Little Games and perhaps his recent post on game analysis: Analyzing a Dataset of Game Releases (or @David Gathercole's Analysis of 'Magic the Gathering' card game ).

On a general note, the whole thread reminded me of the game Heavy Rain. I liked the presence of sophisticated narrative, - more sophisticated than in an average game. I enjoy an original fiction book or movie. I'd love to see that incorporated in games more tightly via interactivity and a good story (NY Times Book Review grade). But Heavy Rain also would progress to a number of different endings depending on the sum of the player's performance even if all the characters become incapacitated in some manner, that's interesting. I am probably too old school and missing a lot of new developments in the game industry.

Again, great conversation !

POSTED BY: Vitaliy Kaurov
Posted 9 years ago

Thanks, Vitaliy. I put the original post together via remote desktop over a mediocre WiFi connection while hanging out with a friend during their bar tending shift. After the connection dropped several times, I gave up on adding any new visuals. Now that I'm back on a good connection, I added one more image of the property influence network for the current full set of rules.

Thanks for the links. They are all interesting. I haven't played Heavy Rain, but I've heard people talk about it. Yes, I would say the goal is to get branching behavior of at least that granularity, but without having to hand script all of the possible outcomes. Perhaps an even more recent step in that direction was Until Dawn. Their marketing even refers to the engine supporting the "Butterfly Effect". I haven't played it though, so I can't personally comment on how intricate or responsive the game world is.

I'm hoping that this design process will identify the most interesting choices and branches a story could take early in development, to help allocate content creation resources appropriately. Probably the last big action game I really enjoyed was Bioshock Infinite. Ironically, for a story about intertwining character arcs in the quantum mechanical multiverse, the game was extremely linear. There were many cutscenes where I thought it would be interesting to make a different choice. Perhaps an approach like this could help developers identify and explore those options early on in development before they are already committed to a particular story or set of choices.

POSTED BY: Michael Hale

Lovely post Michael!

A certain immediacy and intuition seems to be prerequisite in games made today. Why would the average customer choose high investment entertainment over a lower one? I think this has done a lot of damage to the strategy games I used to enjoy. The development of intuitive games falls easily upon realism. If a game looks and acts like the world we know, we discard the investment of learning its rules. This raises the minimum development cost across the board, striving for realism in physics and graphics. This also pushes out the interesting worlds, a grimy war zone more familiar than even a historical setting.

Emergent programming presents a nice balance. The designer can set up a system capable of more scenarios than they've explicitly considered. Though we do find difficult (and more mathematical) problems very quickly! You mention Dwarf Fortress, which is quite happy to let systems interact unchecked. Both the design and the agents in the game face problems from the lack of predictability.

How does one program intelligent characters for unforeseen situations? It's quite easy to establish an oscillation in discrete time systems like these games. How can a player sympathize with a character surprised by longstanding routine events? How can the computational cost of forecasting the future reflect the randomness or regularity a character has actually observed?

Further, how does one control gameplay once off piste? In a scripted game the pace of events and challenges can be mapped out at the designers leisure. Allowing emergent challenges could result in unwelcome fluctuations, long periods of boredom sparsely populated with moments of literal impossibility.

Please keep us up to date as you build the mind of the fisherman!

David

POSTED BY: David Gathercole
Posted 9 years ago

Thanks, David. You raise some interesting questions and thoughts.

I think the amount of time someone will invest in learning a game depends mostly on the entertainment value but then also on potential social rewards for culturally significant games. Entertainment value needs to begin almost immediately and continue throughout the learning curve as opposed to only becoming entertaining after a significant learning investment. The problem I'm running into personally these days, is that the entertainment value falls off too quickly for most games. I somewhat arbitrarily decided that if I get 30 minutes to an hour of entertainment per dollar then I will look back on a game as a good purchase. The problem I have is mostly because the big budget games today are still very similar to the ones I bought and played several years ago. Regarding culturally significant games, I don't actually find chess, for example, to be very fun. It just seems like a somewhat arbitrary example from the class of capture board games with pieces that have varying rules for movement. No one would just sit down and intuitively guess the rules because they resemble the real world. However, it has historical cultural significance. People associate it with intelligence, so I think it is worthwhile to be familiar with the rules. Many popular sports also have a fair bit of arbitrariness in the rules. Also, I suspect most of the dramatic and exciting situations in popular sports from the past few years have happened several times before. However, large amounts of time and money are invested in them due to the potential rewards from their cultural significance.

I think those factors definitely work against new, complex strategy games becoming very popular without significant marketing. I think another factor driving the investment focus on increasing polygon, texture, and lighting detail and rough Newtonian physics for the AAA titles these days is a desire to incorporate scenes that are pulled straight out of popular action movies. That's why so many games have barrels that explode when they are shot.

On the flip side of the realism drive though, I think fantasy stories and settings are more popular than ever in film and television. Look at the success of The Lord of the Rings films or Game of Thrones or of course the new Star Wars movie coming out.

I agree completely that if you are designing a game to be unpredictable, then the amount of enjoyment a player will have could be unpredictable. That's why I'm focusing on adding rules that have many downstream influences. That way hopefully no matter what choice a player makes, the system will respond in some way, which will hopefully keep them interested. Ultimately though, I think I just have to play it as I make it and add details to situations that seem fun as they are discovered.

I think the search approach will handle unforeseen situations naturally. However, I think the challenge will be to keep the computation times feasible as the system grows. If I start adding graphics and animations to a game, it will probably be important to decouple the simulation used for the logic from the simulation used for the presentation. Maybe certain search hints can be pre-computed, or appropriate hierarchical levels of abstraction regarding decision making can be introduced to reduce a large search to multiple small searches.

POSTED BY: Michael Hale
Reply to this discussion
Community posts can be styled and formatted using the Markdown syntax.
Reply Preview
Attachments
Remove
or Discard

Group Abstract Group Abstract