As we enter the Post-COVID world, impending economic downturn takes hold in the business world. Organizations that were performing well well under normal conditions when black swan events like Covid disrupt the normal flow of things. In these scenarios organisations need a system approach to mitigating risk and uncertainty.

As AI researchers we asked ourselves: Can AI help with building strategic resilience to tackle the “unknown unknowns” and give organizations a competitive advantage? How can we help business come up with Anti-fragile processes designs (ones that are not only robust but improve overtime)? and lastly, how can we come up with intervention mechanisms that ensure an appropriate systems response.

In this work we take a systems thinking perspective. we begin by asking ourselves, what can we learn from some of the biggest accomplishments of AI like AlphaGo and then AlphaZero and StarCraft [2]. Such approaches used not only machine learning but went a step ahead by combining classical adversarial game-tree search approaches. We use these as a starting point because we know that most businesses operate in environments which are adversarial in nature.

Suppose we are a business that provides high speed internet connection to 1 million users and Our aim to expand out market share. To do so lets say we come up with a strategy where we promote our services by offering a high-speed internet connection at discount rate. Now suppose we have a competitor with same business goals. There is is nothing stopping from our competitor from copying this strategy. And we counter this by doing an IPO for example and raising funds to continue offering services at a discounted rate. And then our competitor can counter it by doing mergers and various acquisition. You get the idea that this can go forever.

This sort of reasoning about about chains of moves and counter-moves where you anticipate the moves of your adversaries is done a lot in business settings. What business need are way to come up with strategic plans that are both robust and resilient. The approach we propose is of viewing this sort of reasoning and decision making as a 2 player turn making taking game, just like the game of chess or go. We note that in real world its not the case that business and adversarial entities are taking turns one after the other. But we believe that this approach will allow business to efficiently explore the consequences of various decisions in a much more structured manner.

But we need more than just search and learning. We look towards good old fashioned AI of logic and Boolean Satisfiability (SAT) problems. They will provide a mechanism for us to compare strategies and come up with a strategic plan that is both robust and resilient to the actions of the environment.

I’ll start by summarizing our approach and then I’ll explain each step in greater details in the next section.

  1. We start by developing an appropriate goal model for the organisation
  2. Then using this goal model we develop a capability model of the organisation which tells us what it can or cannot accomplish in the given state - We do the same for the environment agent.
  3. We then identify the current strategic decision that we need to make in terms of pursuing a certain sub-goal of our goal model.
  4. We encode the current state of business and the environment its operating in using assertions in the form of of truth-functional assertions and condition-actions rules representing the current capabilities(of the business).
  5. We then select a particular game search algorithm. e.g Monte-Carlo Search
  6. We perform the game tree search to arrive at the decision.

Let's dig in and explore these steps in bit more detail:

Strategy Mapping via Goal Modelling:

Orgnaizational strategies can be effectively captured by goal models. They provide a hierarchic representation of strategic intent, with goals higher in the hierarchy (parent goals) related to goals lower in the hierarchy (sub-goals) via AND- or OR-refinement links.

An AND-refinement of a goal is a statement of know-how that tells the organization how to achieve a parent goal. OR-refinements offer alternative specifications of know-how for a given parent goal. The decision problem we seek to support is that of deciding which of multiple competing goal refinements a business should seek to implement.

Augmented Game-Tree:

Our aim is to develop a machinery that give a given organizational strategic plan resilient to the actions of competitor agents

So we need more than goal models…

To solve this problem we propose a data structure which we call an augmented game tree. An augmented game tree consists of action (or game move) nodes as well as state nodes. It can be generated dynamically at runtime by using the description of current state and capability models of both players.

The notion of state update drawn from the literature on reasoning about action underpins the generation of child states given a parent state and an action that is performed in that state.

Capability models

Actions are represented by their post conditions

Next we need a method for capturing the capabilities for both the business and adversarial entities. In our case this is done by using condition action rules. We use these rules to compute which of the actions are available to us in a given state

We Note that two actions that have the same effect will be seen as actions.

State Update:

Next question is: what will the be the result of a certain action taken by the business. In our case that is determined by the state update operator:

This shows there will be many possible non-deterministic outcomes of a action. The expression tell us that in the resulting state the effects of the last action will be there along with a state set that is consistent with the effects of the action we took and the knowledge base. Resulting set would be the Effects of the current action combined with maximal subset of current state s that is consistent with the effect of the action.

Evaluation Function:

For a game that has a high branching factor and explores the full game tree is computationally intractable. Thus we need a heuristic for exploring the tree until a certain level depending on the compute budget. In our case its done by using an evaluation function which assesses the goodness of a state or a puesdo-leaf node in the game tree. In particular, we use the hamming distance and with the intuition that state which are closest to the current state in terms of change ̄are worth exploring. we then engage in game tree search through the space of states and actions (and not the space of strategies and counter-strategies).

Conclusion:

We offer the means of taking resilient decisions on how to realize organizational strategies by taking into account the organization's own understanding of the limits on its capabilities (as encoded in its goal model) and the capabilities of adversarial entities in its business environment.

References:

1. Berliner, H.J.: Backgammon computer program beats world champion. Artificial Intelligence 14(2), 205–220 (1980)

2. Hassabis, D.: Alphago: using machine learning to master the ancient game of go. Google. https://blog. google/topics/machine-learning/alphago-machine-learninggame-go (2016)

3. Wang, J., Han, J.: BIDE: Efficient Mining of Frequent Closed Sequences. In: Proceedings of the 20th International Conference on Data Engineering. pp. 79–90 (2004). https://doi.org/10.1109/ICDE.2004.1319986