Memory for goals – An activation-based model

When people plan, react to emergencies or other interruptions or make mental notes to do things in the future, an important process involved is the cognitive management of goals.

Very often, a goal must be suspended or set aside temporarily and then resumed later. Suspending a goal might be necessary, for example, if it requires sub-goals to be achieved first, as in hierarchical problem solving and means-ends analysis.

Progress on a goal might be blocked by the state of the world, so one may benefit from turning attention to some other task until the environment changes.

One might be interrupted in the middle of a task prompting the formulation of an intention to resume the task later.

The better the system can remember how far it had progressed toward achieving the pending goal, the more accurately and efficiently it can resume the goal.

1. The interference level is a mental “clutter” of residual memory for goals, which is simply an instance of interference generally.

2. The strengthening constraint says that the activation of a new goal must be increased to overcome proactive interference. This strengthening process is not instantaneous and therefore predicts a behavioural time lag to encode a new goal. Importantly, this lag is functionally bounded from above, because excessive strengthening raises the interference level by making the goal more likely to capture behaviour in the future, when it is no longer the target.

3. The priming constraint says that a suspended (pending) goal can only be retrieved with help from priming from some associated cue, to overcome retroactive interference from intervening goals. The priming constraint has implications for the structure of the task environment, in that the cue to which a goal is associatively linked must be available both when the goal is suspended (in order to create the link) and when the goal is resumed (in order to prime the target).

• Goal directed cognition is often discussed in terms of cognitive structures that happen to mirror the goal-related requirements of the task environment.

  • For example: When a task is hierarchical and requires goal decomposition, the system’s goals are often assumed to reside in a stack, a data structure that provides an appropriate kind of last-in, first-out access. The argument is typically a functional one – because the system can perform hierarchical tasks, it must contain data structures that support the necessary access patterns for goals.

• Two distinctions in early cognitive research:

  1. Control flow – the sequencing of physical or mental steps – and the mental storage structures that support it.
  • If a task is too complex to perform directly, means-ends analysis is a control structure that can be used to decompose it into sub-goals. If these sub-goals are also too complex to achieve directly, they must be decomposed in turn.
  2. Task vs. cognitive constraints.

  • Both kinds of constraint must be met by a system performing a task, but the meeting of task constraints is more easily observed – the analyst or observer need only monitor whether the system’s performance is successful.

  • The influence of cognitive constraints can only be observed indirectly, for example, with additive factors logic. The successful task performance itself proves little about such internal constraints beyond demonstrating the sufficiency of whatever underlying processes are at work.

• The working memory capacity of the system may be less than ideal for solving complex problems and yet the system can manage if it can bring processes to bear that compensate for inadequate (internal) memory.

• Early studies of higher-level cognition and intelligence were focused on the flow of behaviour in response to task structure than on human memory and its constraints.

• One structure well suited to hierarchical control and means-ends analysis is the stack. In a stack elements are ordered by age, with the newest at the top (the most accessible position) and the oldest at the bottom (the least accessible position).

• In cognitive architectures that incorporate a goal stack, the newest goal is the one that directs behaviour. If this goal must be decomposed, because it is too large to achieve directly, its sub-goals are “pushed” on the stack, effectively suspending the goal and shifting control to one of its sub-goals. When a sub-goal is achieved, it is “popped” and the system returns control to the super-goal. Thus, the goal stack delivers the right goal at the right time during means-ends analysis.

• The original analysis of hierarchical plans assumed that all goals relevant to the current activity were also available to direct behaviour.

  • Abstract goals (e.g., being a good citizen) were taken to direct actions (e.g., giving alms to the poor) concurrently with specific goals (e.g., helping a particular person).

• Special goal structures become a more explicit assumption with the development of the ACT and Soar cognitive architectures.

  • ACT – goals were sources of activation without requiring active maintenance and were linked associatively to other goals to which they were related by task constraints. This meant that returning control to a super-goal was a reliable operation since the sub-goal simply pointed to it.

  • Soar contains no notion of activation – items are either in working memory (WM) or not – but the goal stack is part of WM and hence its contents are always immediately available.

• The stack predicts a pervasive tendency to select goals in a last-in, first-out (LIFO) order, but goal-selection order is highly idiosyncratic in non-LIFO tasks and the stack is difficult to “patch” in a way that explains systematic forgetting of certain types of goals.

• Behaviour that seems to be governed by serial goal selection can sometimes be attributed to perceptual-motor constraints and where a LIFO goal ordering is necessary; it can often be reconstructed from the environment.

• Existence of active inhibitory processes:

  • Intentions are stored in a special state, active inhibition is required to remove them from working memory.

  • The intention superiority effect. The finding is that memory for an action script is better than memory for a “neutral” script with no associated action and also better than an “observe” script where the action is to be observed rather than carried out.

  • Goal nodes are the only elements of working memory in ACT that do not need rehearsal to sustain activation.

  • The goal stack was a response to the original challenges of cognitivism – it was necessary to first trace complex mental behaviour, for which the stack was eminently useful, before delving into its relationship to lower-level constraints.

• ACT-R – derives important constraints from asking what cognitive processes are adaptive given the statistical structure of the environment.

• In human memory, old items decay gradually rather than instantaneously, forming a mental clutter that can make it difficult for the system to find the correct item.

• Informally, the interference level represents mental clutter.

• Formally, the interference level is the expected (mean) activation of the most active distractor.

• When the system samples, it will get the most active goal. Whether this is the target or the most active distractor depends on their relative activation values.

  • If the target is above the interference level, then it is more likely to be sampled than any distractor.

  • If the target is below the interference level, then it is less likely to be sampled than some distractor.

• As long as a goal is above the interference level, it directs behaviour.

• Activation is subject to noise.

• A new goal will suffer proactive interference from old goals, so to direct behaviour it has to become more active than the interference level.

• The more active an item is now, the higher the interference level later when that item is not the target.

• The system must strike a balance between making a new goal strong enough to direct behaviour now, but not so strong that it interferes excessively later once it is satisfied.

• The functional constraints have behavioural implications in that they help predict the amount of time that the system needs to spend encoding a new goal.

• During strengthening, the system devotes all its cycles to sampling, causing a rapid build up of activation. Then the system turns to other activities, using the goal for direction. Such activities might include formulating another goal, or might involve manipulating the task environment in a goal-directed manner.

• Strengthening is serial – one sampling event occurs at a time.

  • Activation increases with exposure to a stimulus.

  • Strengthening time is bounded from below by the time needed to bring the new goal above the interference threshold.

• Strengthening is cumulative – its effect persists throughout the goals lifetime.

  • Although the system could improve its memory for a particular goal without limit by spending more time on strengthening, life is not so simple that we have a single goal.

  • Goals often arise every few seconds, so the interference level is a critical limit on how much activation the system can afford to invest in any particular one.

• The strengthening process makes testable predictions through its role in planning – a process of mental stimulation in which the system imagines a sequence of problem states starting with the configuration shown in the task environment.

• The strengthening process is what makes the state immediately and reliably available, by making the state more active than its competitors.

• Strengthening extends in time, but this time is bounded because the system needs to avoid too much of a good thing.

• Planning often incorporates backtracking – returning to an immediate state in the problem space after pursuing other paths.

• A suspended goal suffers retroactive interference – it will be “buried” or “masked” by its successors in terms of activation – so to direct behaviour it has to be made more active than the interference level.

• Activation can be analysed in terms of history and context – the assumption is that the memory system integrates prior experience with current evidence to predict current need for a memory element.

• Cues provide associative activation, or priming, to a target item to which it is associatively linked.

• Priming is the only way to overcome the retroactive interference that affects an old goal. “Now what was I doing?” can only be answered by generating the right cue as a reminder.

• The priming constraint has an important implication for cue availability.

  • The system must attend to the cue when trying to resume a target goal.

  • The cue must be associatively linked to the target goal in the first place.

• In ACT-R, links between cue and target are often formed by co-occurrence – the cue and target have to enter the system’s focal awareness at roughly the same time.

  • The implication is that the goal and its retrieval cue must have been sampled together, before the goal was suspended.

• When the task is such that the solution path cannot be memorised, the task environment must contain means-ends cues if old goals are to be retrievable.

  • They will be “obvious” in some sense – obvious enough ideally that it would be difficult for the system not to process them.

  • Means-ends cues will prime the target but relatively few distractors.

• Retrieval cues need not be environmental – they can be drawn from long-term knowledge about the task and thus be part of the problem solver’s internal mental context.

• When the cognitive system asks, “What am I doing?” it is in effect sampling memory for a goal. The sampling is complicated by the interference level, which represents the activation of the most active goal that is not the target.

• To direct behaviour, a new goal must be strengthened to overcome the interference level and a suspended goal must be primed by a retrieval cue to overcome the interference level.

• The strengthening and priming constraints make predictions both for behaviour and for the structure of the environment.

• The strengthening constraint predicts that setting a new goal takes time (evident though the need to plan ahead).

• The priming constraint predicts that cues must be available that the system can encode with a goal and use to prime retrieval later.

• Means-ends analysis is sufficiently important to human behaviour that it might be viewed as the opposable thumb of cognition.

• Means-ends analysis allows us to grasp and manipulate complex concepts piecewise when the whole is too difficult to manage.

• Strengthening each goal is critical during planning moves.

  • Each goal is the basis for formulating its sub-goal and although the initial goal on such moves can be formulated directly from perceptual information, its sub-goals are intermediate mental products and have only mental representations. Because a sub-goal resides in the head only, it must be active enough to be a reliable input to the next level of recursion. The cost associated with an insufficiently active sub-goal is that a memory failure breaks the chain of inference and forces the system to start over.

  • Strengthening goals supports their retrieval later. The encoded goals represent a plan and it pays to be able to retrieve the steps of this plan. Retrieving a plan step requires not only a retrieval cue, but also depends on the strength of other goals primed by that cue.

• The priming constraint says that retrieval cues are necessary to resume an old goal because the old goal is affected by retroactive interference from intervening goals and needs the activation boost.

• The prediction for the Towers of Hanoi is that for every goal retrieved, there is a means-ends cue – and object likely to be attended both when the goal is suspended and when it is to be retrieved and that primes that goal but no (or few) others.

• Errors are caused by activation noise, which can make the wrong goal strongest and send the simulation down the wrong path.

• The more the simulation relies on memory for goals, the more opportunities it will have to retrieve the wrong goal and stray off the optimal path.

• Errors should increase with problem size, as larger problems involve greater reliance on memory.

• The most active goal in memory is the one that directs behaviour, because that is what the system samples when it seeks guidance from memory.

• During planning, as the system decomposes a large goal into smaller goals that are easier to achieve, the goal being decomposed must be highly active in order to be the reference point for the decomposition process.

• Later, during execution, the same goal may have to be activated again once its sub-goals have been achieved. At this point, priming for cues is necessary to overcome retroactive interference from other, newer goals in memory.

• Environmental cues are necessary and sufficient for goal reconstruction in dynamic environments.

• There’s a strong dependence of skilled or expert behaviour on environmental cues.

• Progressive deepening – as goal states become more complex and cue-selection heuristics less obvious, we would expect problem solvers to plan from the external state more often, as the difficulty of retrieving intermediate states from memory increases.

• Progressive deepening is common in conceptual, ill-structured domains.

• Despite the advantage that recency generally confers on memory for items, inhibition of return is found in diverse cognitive behaviours, from task switching, to visual attention, language processing and implicit sequence learning.

• An intention to perform an action (an action intention) is stored with greater activation in memory than an intention to observe another actor doing the action (an observe intention).

• Memory for something is better the more “goal-like” that thing is.

• Intention superiority could simply reflect strategic memory processing taking place in the temporal chinks of the procedure.

• A second or two of strengthening can have substantial effects on memory-based performance.

• This is a memory-based error made whilst “wrapping up” some common procedural activity. Examples include:

  • Take the copies but forget the originals.

  • Take the cash, but for get one’s cared in the cash machine.

  • Drive away from the petrol station, leaving the car’s petrol cap on the roof of the car.

• The forgotten action occurs after the main goal of the activity is accomplished.

• Post-completion actions are goals themselves in the sense that one “knows” to carry them out as part of one’s procedural knowledge.

• The user’s procedural knowledge includes an associative link from the main goal to the actions necessary to achieve it. This link spreads activation to the actions as long as the main goal itself is active.

  • The main goal of getting petrol primes the action of replacing the car’s petrol cap and thereby keeps the action available as a pending sub-goal – as long as the tank is not yet filled.

• An assumption is that correct performance depends on the task taking up most of the available working memory capacity.

• Activation moves around among memory elements but the overall quantity is conserved.

• Insufficient memory load raises the interference level.

• Interference cannot be the only kind of forgetting – interfering elements themselves must decay if memory is to serve its function.

• The decay process in the goal-activation model has a cost as well, which is that suspended goals are forgotten gradually, making them harder to resume.

• The general implication is that people are able to retrieve suspended goals successfully if and only if there are cues that meet the priming constraint.

• Interruptions are continual – this creates a situation in which a goal must be suspended before it is completed and then resumed later.

  • The first even (the phone ringing; an alarm sounding) can be viewed as the alert.

  • The subsequent event can be viewed as the interruption proper.

  • The time between onset of the alert and onset of the interruption is the interruption lag.

• Efficient resumption depends on associative links between environmental cues and the target goal. Such links must be formed before interruption onset (assuming that the interruption truly draws the system’s attention away from the interrupted task). The interruption lag is a natural window of opportunity to form such links.

  • By analogy, if two people are conversing an the phone rings, the callee faces the strategic decision – whether to use the available time while the phone is ringing to end the meeting with a few final words, or to reschedule it to continue at a later time.

• The processing of cues during the interruption lag may be relatively automatic, in which case simple manipulations of duration of interruption lag or the availability of cues should affect resumption efficiency. On the other hand, such processing may be deliberate, if it occurs at all, suggesting that it could be trained.

• Goal-directed behaviour can be explained with the general memory mechanisms of activation and associative priming.

• Activation initially comes from strengthening (or encoding, or “paying attention” to) the target goal and is necessary to keep a goal active during any kind of planning or other mental stimulation.

• Later, if the goal has been suspended and the time comes to resume it, associative priming is necessary to retrieve it from memory because it will be less active than the goals encoded in the interim.

• Associative activation comes from cues linked to the target goal; cues that may lie in the environment or that may lie in long-term mental representations like procedural knowledge.

• Cues must be associatively linked to the target goal to be of any use as primes. Cue availability and selection are important factors at goal-encoding time as well as at goal-retrieval time.

• Altmann, E. M., & Trafton, J. G. (2002). Memory for goals: An activation‐based model. Cognitive science, 26(1), 39-83. https://doi.org/10.1207/s15516709cog2601_2