Data Scientist at Sensum Co., Founder/Director at Farset Labs
It’s nearly a year to the day since I passed my Ph.D Viva (And since I last updated the blog…), so I thought it’d be fun to very-gently tidy up one of my appendices that’s a bit relevant to current stories about the end of the world and machines taking over and such.
It’s a piece of work that I enjoyed researching and had originally had as a significant part of the main thesis, but it just didn’t fit in anywhere sensible, so it got stripped to it’s bare minimum and kicked to the end.
If there’s an interest in this stuff, I’ll have a go at revisiting the most painful document of my career and tear out the interesting bits!
So without further ado, here’s a rant about how human factors and HCI impact human/AI teams.
This work has largely considered autonomous systems as entities of wider systems, implicitly involving human operators/agents in some part of the desired operation. We refer to these systems as Autonomous Collaborative Systems (ACS). As described in Chapter 2 Operational Trust has two main aspects, trust in the system to behave as expected and trust in the interfaces between systems (human/machine and machine/machine). Of all of the interfaces in an Autonomous Collaborative System, the most problematic is that arguably that between the ACS and the human operator / team of operators. Cummings et. al. (2010) identified the main challenges to Human Supervisory Control (HSC), summarised below:
Operator efficiency exhibits an optimum at moderate levels of cognitive engagement, above which cognitive ability is overloaded and performance drops (Otherwise known as the Yerkes-Dodson Law). Additionally, in the case of under-engagement, operators can fall foul of boredom, and become desensitised to changing factors. However, predicting this point of over-saturation is an open psychophysiological research problem.
Automation is well tailored to consistent levels of activity. This is quite simply not the case in many domains. Particularly in defence and military applications, activity is characterised by long periods of “routine” punctuated by high intensity, usually unpredictable, activity. At those interfaces between “calm” and “storm”, where real time situational awareness is imperative, temporary Information Overload is highly probable. Adaptive Automation enables autonomous systems to increase their Level of Automation (LOA) based on specific events in the task environment, changes in operator performance or task loading, or physiological methods. It is taken as given that for routine operations, and increased LOA reduces operator workload, and vice versa. However, this relationship is highly task dependent and can create severe problems in cases of LOA being greater, or indeed lesser, than is required. In the cases of overly-high LOA, operator skill is degraded, situational awareness is reduced as the operator is not as engaged, and the automated system may not be able to handle unexpected events, requiring the operator to take over, which, given the previous points, is a difficult prospect. Alternatively, in sub-optimal LOA, Information Overload can result in the case of high intensity situations, but also the system can fall foul of overly-sensitive human cognitive biases, false positive pattern detection, boredom, and complacency in the case where less is going on. Therefore, as a corollary to Information Overload challenges, there is a need to define the interrelationship between levels of situational activity (or risk) and appropriate levels of automation. Under what circumstances can adaptive automation be used to change the LOA of a system? Does the autonomous system or the human decide to change LOA ? What LOAs are appropriate for what circumstances?
Distributed Decision Making
In a modern, non-hierarchical, often distributed or cellular military management system (Network Centric Warfare doctrine for example), tools are increasingly being used to mitigate information asymmetry within command and control. A simple example of this is shared watch-logs in Naval operations, providing temporal collaboration between watch-teams separated in time. The DoD Global Information Grid is another example of a spatial collaborative framework. Recent work has demonstrated the power of collaborative analysis and human-machine shared sensing technologies even with low levels of training on the part of the operators providing superior results and resource efficiencies than either humans or machines alone in survey and search-and-rescue scenarios (Ahmed et. al. (2014)). As these temporal and spatial collaboration tools increase in complexity and ability, decisions that previously required situational awareness that was only available at higher echelons within the standard hierarchy are available to commanders on the ground, or even to individual team members, enabling the potential for informed decisions to be taken faster and more effectively, enabled by automated strategies to present relevant information to teams based on the operational context. However there are a range of operational, legal, psychological and technical challenges that need to be addressed before confidence in these distributed management structures can be established. Studies into situational awareness sharing techniques (tele-present table-top environments, video conferencing, and interactive whiteboards) have generally yielded positive results, however investigations into interruptive-communications (such as instant messaging chat) have demonstrated a negative impact on operational efficiency. In short, the biggest problem with distributed decision making in the context of supervisory systems is that there is no consensus on whether it is advantageous or not, and what magnitude of operational delta is introduced, if any.
Beyond simple Information Overload, increasing complexity of information presented to operators is having a negative effect on operational efficiency. In HSC, displays are designed to reduce complexity, introducing abstractions with an aim to presenting the minimum amount of information to the operator required to maintain an accurate and up-to-date mental model of the environmental and operational state. This has led to the development of many domain specific decision support interfaces, however, in academic research, there has been nothing but ‘mixed results’. One commonly raised negative is the general bias on the “cool factor” of interfaces. Immersive 3D visual, aural, or haptic interfaces that at first appraisal seem to provide more approachable information to the operator, and are indeed tacitly preferred by operators in use. However, there has not been any evidence to demonstrate performance improvement when using these tools, and in-fact, improving the “fidelity” of the interfaces has led to operators’ overly-relying on these representations of the environment rather than remaining engaged in the environment.
Cognitive Biases and Failing Heuristics
In many areas, operators are required to make rapid decisions with imperfect information, driven by massively increased information availability and rates of change in areas such as battlefield tactics and global finance markets. However, Human decision making isn’t always rational (especially under pressure), and operators use personally derived heuristics to make “rational shortcuts”. This is a double edged sword, where these heuristics can be employed to greatly reduce the normative cognitive load in a stressful situation, but also introduce destructive biases, where these shortcuts make assumptions that don’t bear out in reality.
For example, in the context of decision support systems, “Autonomy Bias” has been observed as a complement to the already well known “Confirmation Bias”1 and “Assimilation Bias”2, where operators that have been provided with a “correct” answer by a decision support system do not look (or see, depending on perspective) for any contradictory information, and will unquestionably follow, increasing error rates significantly.
This behaviour isn’t only the reserve of decision support systems, but also in the generic allocation of operator attention; scheduling heuristics are used to decide how much time tasks should be worked on, and time and again, humans are found to be far from optimal in this regard, especially in time-pressured scenarios where these heuristics are in even more demand. Even when operators are given optimal scheduling rules, these quickly fall apart, often due to primary task efficiency degradation after interruption. This highlights a critical interface in the adoption of complex autonomous systems that still demand ‘Man in the loop’ functionality; if a system is required to have full-time concentrated supervision (e.g. flying a UCAV), but also event-based reactive decision making (e.g. alerts from non-critical subsystems), both tasks are negatively impacted. In an assessment of factors influencing trustutono in autonomous vehicles and medical diagnosis support systems, Carlson et al also identified that a major factor in an operator or users’ trust in a system was not only dependant on past performance and current accuracy but also on “soft factors” such as the branding and reputation of the manufacture / designer (Carlson et. al. (2014)). Further, autonomous decision support / detection / classification systems have an “uncanny valley” to overcome in terms of accuracy, in that there is a dangerous period when such systems are used but not perfect, but operators become complacent, causing an increased error rate, until such a time that those autonomous systems can match or exceed the detection rates of their human counterparts.
The separate fields of automation and user experience design have been running in parallel for several years. However, there will soon come a point (in some cases already past) where human operators must place their “Trust” in autonomous systems to not only accomplish what they want and what they expect, but to do it in a way they are psychologically comfortable with. Further, there is the aspect of how are (or should) autonomous systems being “trained” in how to deal with the systematic failings in human cognition? At what point does the machine need to trust the operator before it performs “responsibly” in the face of a possibly irrational or cognitively broken usage, and if so, how can it communicate this state to the operator or some higher order command structure?
These are massively open research questions I didn’t get to attempt to answer anywhere other than the appendices and the pub, and there is no easy avenue to start from, so this author suggests Asimov’s “The Robot Collection”.
Confirmation Bias is the tendency for people to preferentially select from available information that information that supports pre-existing beliefs or hypotheses. ↩
Assimilation Bias is often thought of as a subset of Confirmation Bias, whereby it specifies that instead of seeking out information supporting of current views, any incoming data is interpreted as being supportive of a particular view without questioning that view, even if it appears contradictory. ↩