Reading Between the Lines: Learning by Experience

Feb 28, 2011

This graphic represents a series of decisions, and the success of acting on those decisions, each of which leads to other opportunities to make decisions and act on them. As time progresses, the player advances in circumstances along the arrows in the direction they are pointed. Taking one arrow rather than another at a forked intersection might mean picking up one item instead of another, taking one approach to an encounter rather than another, making one puzzle adjustment rather than another, or failing rather than succeeded to execute an attempted task.

Not (Necessarily) Tied to Space

To be clear: this is not a map of navigable space, though a split in the flow could map to spatial decision or action result, such as taking a ladder onto a dock instead of swimming under, or falling down a cliff after failing to catch a ledge. The actual space of decisions and action consequences is likely far larger and more dense than what is illustrated here; think of the graph as a tiny section of a relatively simple example, not as a complete representation capturing the depth of a general case.

The Experience Highlighted

In a hypothetical series of decisions and action results colored above, the player first traces the red line over time, from A to B, ending up in a failure state (marked here by skull and crossbones). Upon restart, the player begins by making a different decision, following the green path of consequences from A to C, and has a victory confirmed by the program (the circle around a dot inside at C).

Based on the experience described, the player now has personal knowledge of two possible sequences through the game, one which led to failure and one which led to success. What does that player really know about the game at this point? To answer that question, let us explore how the knowledge gleaned from that experience might inform future play attempts.

Informing Future Experience

Were our initial player to witness another player getting into a problem, or even to simply catch themselves during a future play session getting into a problem, a warning (another player) or internal thought (initial player on future attempt) would suggest redirecting play. There are two forms that our initial player might recognize as “getting into a problem”:

  • By State: The logical circumstance at the moment is identified to be part of a failed path. In the graph above, this would be mapped as a point along any red highlighted arrow.
  • By Decision/Action Process:The rationale being used to make decisions, or the technique applied to execute actions, resembles that which previously steered a path into a failure condition (dead end, or alternatively and nearly equivalently, an unproductive loop).

The response to the issue might be to backtrack (if viable), to quit and restart (depending on the amount of progress lost in doing so), or going into a hypothesis-action loop until no longer in a state or applying a decision heuristic/algorithm warranting concern.

Colloquially, these two possibilities translate to, “You’re in trouble, when I was in that situation it led to failure,” (State), and, “Don’t go about doing it in that way, because I did and I failed,” (Decision/Action Process). Thinking about them in those terms reveals the problem with assuming that either information type alone is sufficient evidence of certain concern. Someone could make different decisions from a similar circumstance, or similar decisions from a different circumstance, and arrive at a state different than what happens from making similar decisions from similar circumstance. (Note that randomization of luck or difference in skill for executing action may guide the selection at a future fork in a different direction, but for sake of coherence, those factors are assumed consistent.)

Blindness to Potential Experience

Note the optimal path marked in blue, branching from halfway along the red arrows at D. Compared to the green path, all we can see from this depiction is that it is 1 decision simpler, though it might also or instead be better by some other criteria not shown, for example by involving less unpredictability, or requiring less skill/practice.

Our initial player, based on the red line and green line experiences traced, may be inclined to warn others (including his future self) in such a way that would prevent discovery of it, and could only find it by retreading steps that previously led to failure.

(Repeat of the first illustration, to minimize scrolling.)

The Experiences Not Experienced

What, if anything, is known by our initial player about the yellow decision/action paths, of which nothing has been seen?

Whether so constrained as a sequence of multiple choice questions, each possibility leading to more multiple choice questions (like Masq), or so fluid as deathwatch in Quake 3 or even more so in a non-digital sport (say, soccer or boxing) – unless the experience exists in a precisely linear, one-way, forced progression like reading a book or watching a film having no room for branching decision or different outcome of action (as seen in all but a few screens of Transcend), these unseen experiences greatly dwarf the number of varieties of experiences actually seen.

Consider the potential arrangements of Tetris blocks that a player has never seen in comparison to those that the player has observed, or the combinatorial explosion of blocks missing/present in a moment of Breakout gameplay. Even in these comparatively old, simple videogame experiences, the states experienced are a tiny fraction of the states possible.

Those two examples help lead the conversation to the next step: clearly, we could show any player of Tetris or Breakout a possible on-screen configuration (these games are convenient since their entire game state is volunteered unambiguously on-screen at once, spare the Breakout ball’s present velocity), or drop them into a game in-progress, and they could proceed without a hitch. Not only do they know something about this experience they have never experienced, they understand it rather exactly, as though it were a situation that they have been in before.

Before interaction has occurred, these grounded assumptions about what to do are based on affordance. That is, to what extent do clues volunteered by an object remind us of things to which we are already familiar? Even if we drop a Breakout player into completely different implementation with different graphics, different initial layout, and different sounds, similarity in object relative location, scale, and movement would volunteer expectations about how these different parts will behave and interact.

After some interaction has occurred, these grounded assumptions are informed by expectation of learnable consistency: that what looks like will function alike, that all things will continue to function moments from now in a way strictly consistent with how they functioned moments ago, and that if there is randomness, then the likelihood of something random happening is consistent with the frequency in which it has happened in the past.

(Note that both of these assume, or rather require, perceivability. If there is no clue, indication, or measure offered through visual depiction, motion, audio, changes in control, etc., then a consequence suddenly perceptible to the player based on causes not perceived might be interpreted as random, or perhaps stimulate exploratory observation in an attempt to make sense of what took place.)

This is About Learning

Unless the purpose of a test is to become better at passing tests (such is the case, for example, for an SAT pretest), or to otherwise exactly reproduce a particular result (memorizing digits of pi, accurately pinpointing capitals on a map, etc.), the goal of learning is frequently to prepare the learner for cases not yet exactly encountered. The purpose of a math exam is to discern, through a small but unknown sample set of the possible problems taking a certain form, whether it seems likely a student is able to accurately answer any question of that form.

A skilled Quake 3 player, at any time in the deathwatch, could have the level, weapons loadouts, and character positions scrambled randomly, and continue to perform, because mastery of Quake 3 is more about the decision heuristic/algorithm than the state (except to the degree which reading of state serves as input to or feedback within the decision heuristic/algorithm).

(Repeat of the same illustration, for text references.)

The optimal blue path, in particular that it is hiding behind the red path – especially once the non-optimal but successful green path has been discovered – represents a better way to tackle a problem than what someone will allow themselves learn, try, or recognize, due to negative weight assigned to the entire sequence which has potential diverge early on into the superior strategy. These types of buried paths are a sign of mastery, and they get lost as an effect of due to iterative learning.

Those properties above identified as what makes the yellow, unseen decision states as predictable (affordance, learnable consistency) are more present in some situations than others. Our knowledge of what would result form a single different decision in an interactive fiction game like Masq is much less than our knowledge of what could be assumed in a novel situation or following a different decision in Quake 3. A game like Masq exists as a network of connected states with little/no learnable consistency. There are still clues embedded into the fiction – frankly all Four Aspects – in that we can expect certain things about consistency in the behavior of characters (assumed to be consistent with, or at least reasonable by, modern cultural norms of human behavior), in the arc of the plot (following a certain expectation by its genre), and so on.

A decision heuristic/algorithm can still be employed in a game like Masq, such as always avoiding violence, or always being “a good guy”, but any given decision could lead to a total non-sequitur, in a way that we wouldn’t tolerate for a game like Quake 3 (if we walk toward an ammo box with no one in sight, we expect to get it upon arrival, not to have it vanish when we’re 6 units away – unless, as a learned convention, it was a drop from a fragged player). As soon as “magic” of that sort begins happening in a real-time game, inconsistency in a way that seems unlearnable, an impression is given that skill cannot be gained in the game, or rather than additional practice will not enable a player to more effectively progress or perform in the game.

An impression that the game’s states are not connected by learnable consistency – even if only through coherent fiction construction rather than mechanical continuity – leads to frustration. The states may actually be connected by learnable consistency, but something about how it’s exposed through the Four Aspects gives the player the impression that it isn’t, which hammers the same stake into learning as if it were actually connected without learnable consistency.

I suspect, though cannot show at this time, that a very similar issue arises in learning in general, particularly learning by experience.

More on that soon.

Learn and practice team game development with Gamkedo Club.
Membership worldwide. Professional support. Proven process.

Subscribe by e-mail to receive weekly updates with Gamkedo.Community interviews and YouTube training videos for game developers!

All contents Copyright ©2017 Chris DeLeon.

Site production by Ryan Burrell.