A.I IN AFL PART 1/3

                                                      

 

Last week Sam Mitchell caused quite a stir when he admitted to using AI in his coaching process and although what was reported contained very limited detail of what he actually uses and what he uses it for, the uproar was ridiculous.

There is no doubt that AI will become a part of Aussie Rules in the near future as team's are always looking for ways to optimise and streamline their approaches and there's not an easier way to do that then through AI means.

On Twitter I came across this comment on the subject which of course sent me down the rabbit hole of this 341 page study titled "Artificial Intelligence-Based Decision-Making Support During Australian Football Matches" and here are my numerous notes from it.

• During games coaches face limitations to available information and cognitive capacity (calculating risk/reward)
• A decision support system takes complex computational methods and empowers the user with them in an interactive/easy to use manner
• Can pinpoint in-game performance indicators and their relationship to success in that specific game
• Phase 1 x exploration of in-game decision-making of AFL coaches
• Phase 2 x develop a decision-support systems (DSS) to leverage AI/Performance Indicator’s (PI's) predicated on the insights gained from phase 1 to support coach in-game decision making
• Phase 3 x deployment/implementation of dss with AFL clubs to evaluate DSS impact
• Decision making is intertwined within the coaching process in both participation/performance settings
• Coaching is built on a cyclical relationship between information, knowledge, skills of the coach, athlete's capabilities, performance analysis, competition and preparation, all that require elements of decision making
• The need to make complex decisions to fit the problem at hand may result in some parts of the coaching process receiving more attention than others
• Coaches employ naturalistic decision making, consulting common sense models developed through experience with the premise of decision making being that decisions emerge from a subconscious process of scanning the environment, recognising a decision problem (a mental threshold is triggered) and consulting knowledge stored as mental models without explicit reasoning resulting coaches using mental simulation/consultation of previous matching decisions from decisions in time=pressured situations
• All decision making requires ongoing situational assessment that triggered some form of action which suggests that there is more to coaching than experience/intuition and it relies on the idea that the formulation of decisions still rests heavily on the cognitive processes of coaches
• Coaches should consider their decision-making environment
• In complex situations, individuals trying to make rational decisions (based on reasoning) usually end up making satisfactory ones but not the optimal one because rationality is constantly bound to the interactions of cognitive/environmental constraints = suboptimal/inconsistent decisions with making the right decision being heavily dependent on the information available and how well a coach interprets it
• Cognitive constraints (memory, heuristics, knowledge, problem solving ability) + environmental constraints (sport science, assistant coaches, performance analysis, athlete monitoring technology, available information) = rational decision making capabilities
• Inconsistent decisions will be made based on instinct alone
• Coaches need to consult with both cognitive/environmental constraints to make informed decisions
• Machines can process larger/more complex loads of information than humans and more rapidly
• Human data collection can be inconsistent from different interpretations
• PI’s are quantitative measurements of action variables that aim to define different aspects of performance and they should form the basis of analysis
• Research in AFL found that winning teams were different to losing teams with regards to 14 out of 16 analysed PI’s including greater values for kicks, handballs, marks, inside 50’s, un/contested possessions and tackles
• In AFLW, PI's were disposal efficiency, kicks, marks and uncontested possession = winning
• Coaches can use data to make decisions to maximise/minimise certain PI’s
• Rather than using PI’s to compare un/successful performances, use PI’s to group players into positional roles
• PI’s related to ball winning, ball use, hit outs, defense and negative performance can allow coaches to identify players of a desired positional role while supporting recruitment/selection decisions
• Feedback can be categorised by valence (positive/negative), prescriptive (information on how to change performance)/descriptive (described performance) + content/audience
• Post-game feedback retention from players is low possibly from being ineffectively delivered and it is more often descriptive v prescriptive which allows athletes to self-regulate their response to feedback in a way that is conducive to learning
• Rather than trying to immediately correct performance, the coach is afforded time during the week to consider the best way to correct performance through further feedback/training design
• Pre-competition feedback should be targeted towards decision-making by facilitating knowledge creation so they need to allow players to solve problems independently while guiding them in the right direction
• Training is the time to manipulate the quantity of feedback to develop problem solving abilities with pre-post competition feedback focusing on providing players information they need to form their own decisions
• AFL has a runner to deliver messages with 81.3% being prescriptive and 58.4% were controlling in nature where coaches pressured players to think/act in a certain way = a higher proportion of prescriptive feedback v other sports
• Coaches send out 15 messages/quarter with the highest amount coming at times when the game is undecided and significantly lower in games when the game was "over"
• Expert decision makers recognise cues/patterns quicker than novices but are still limited by the information processing capacity of humans and machine learning can combat these limitations by finding patterns that the coach doesn’t have time to and decreasing the need to analyse raw stats and have more time to apply expert judgement with a higher standard of information than previously
• Decision trees classify outcomes by splitting the data on features that provide the most information about an outcome until splits can no longer be made such as meters gained relative to turnover’s forced score relative to time in possession relative to win/loss + meters gained relative to turnover’s force score relative to win + meters gained relative to turnover’s forced score relative to disposals to win/loss
• Trees terminate the splitting process when a selected information threshold/tree length is reached
• Of 97 PI’s available for analysis, decision tree model accuracies were as high as 88.9% for finding the most important PI’s to winning being meters gained, inside 50’s per shot on goal, turnover’s forced score and time in possession
• Although accurate/easy to interpret, they can be too simple to provide great insight into performance
• Via a generalised linear model, score margin was predicted on average within 7.4pts with the most important measures being inside 50’s per shot, inside 50’s and rebound 50’s
• Converted to win/loss outcomes, generalized linear model sits at 95.1%
• Technical PI’s can explain performance just as good as tactical PI’s but tactical provides a more unique insight into performance
• Cognitive constraints (coach's perception, memory of events, philosophy, opinions) + environmental constraints (live match stats, athlete monitoring technology, information from assistant coaches) + suggestions from decision support system all go to the coach's mind = decision is made and communicate to players
• To make a DSS these question’s needing to be answered include how do coaches make decisions in competition, what barriers to effective decision making do coaches face during games, what are the thoughts/perceptions of coaches towards the use of AI-based DSS’s during  games, what are the visual/functional design elements which would enable coaches to utilise/interpret information provided by a DSS, how efficacious are score agnostic models for the prediction of end-match outcomes in games, which variations and what type of model is the most useful for supporting coaching decisions during games, when implemented in the field what is the impact of an in-match DSS on coach decision making during games, how does the way a DSS is used change throughout an implementation, what are the elements of a DSS or its implementation process which contribute (positively/negatively) to the uptake of the system…
• In-game decision making consists of 6 stages x opportunity trigger, understand the opportunity, determine the need for action, explore options, take action, evaluate
• Bounded rationality, an extension of rational choice, describes the process that coaches employ to make decisions with the outcome of the decision dependent on the interaction between the available information, their own cognitive limitations/biases and the finite time in which they have to act and given the same decision making prob, it explains how different coaches can come to different right decisions
• Coach decision making study - bounded rationality highlights the benefit in having access to the best possible information/environment with masterful mental capabilities (perception, knowledge, metacognition) when making decisions
• A common framework for understanding the decision making of sport coaches is naturalistic decision making and it explains that in time-constrained decision making environments, expert coaches will scan for/attend to key attractors/recognisable patterns in a continuously unfolding environment, framing a decision problem if a mental threshold is reached
• Key attractors immediately lead expert decision makers to a simply matched solution in most cases though they search for additional cues to diagnose the problem if an immediate solution is not recognised
• In some cases, a stored solution may not match the present problem leading to mental simulations of potential modifications to solutions from previous experiences to evaluate their applicability to the present problem
• Coaches have more time to make their decisions in v players/umpires and therefore can engage in more critical reasoning