# Deep Probabilistic Decision Learning Returns Perfect Flow to Operations

*FlowOps enables optimal experience, predictions and decisions in the operations of factories and supply chains.*

In human brains, there are three key learning functions related to how we sense, predict and decide. Findings in computational neuroscience [1, 2] suggest that different parts of brain areas play a distinct but connected role in each function. These can be equated with the three Explainable AI (XAI) engines in Noodle.ai’s FlowOps products: **Sentinel** (Deep Learning), **Precog** (Bayesian Probabilistic Learning) and **Pathfinder** (Reinforcement Learning).

The interplay between deep learning and probabilistic learning are similar to a human brain’s thinking fast and slow like in Kahneman’s System 1 and System 2. System 1 is a fast, intuitive, heuristic, deterministic, differentiable, and more affective mind, whereas System 2 is a slow, deliberate, logical, probabilistic, integrating, and more cognitive mind.

Deep learning (Sentinel) enables fast, scalable, and associative pattern detections from high-dimensional, noisy and temporally correlated data, using differential optimizations on flexible functions with deterministic model parameters. In contrast, probabilistic learning (Precog) allows slow but causal model-based predictions on future observations for a controllable action and an exogenous condition at a current latent state of the dynamic process, using integrations over probabilistic latent variables to incorporate uncertainties. Reinforcement learning (Pathfinder) trains an action policy for optimizing a long-term cost function over future conditions.

That is, **deep learning, probabilistic learning, and reinforcement learning,** respectively, play a key role in FlowOps, equating to Data (state representation) + Model (causal predictive modeling) and Policy (policy optimization).

### What is Flow Operations (FlowOps)?

FlowOperations, a new category of Enterprise AI® software, is well described by Noodle.ai’s CEO in his Pi Day blog for the launch of the category. In it, he also defines Operations Entropy as “… the seeming randomness, unpredictability, uncertainty, and general fog that happens every day in complex factories and distribution networks. Complicating things further, entropy increases as complexity increases.” Our three XAI engines, Sentinel, Precog and Pathfinder, address entropy and are based on the following principles:

In general, FlowOps aims at reducing the flow (value)-at-risk or any unnecessary cost due to bad operations flows. There are a few key information-theoretic entropy principles we apply to build our FlowOps.

In **Sentinel**, we maximize the evidence lower bound of our model (or called the negative variational free energy) to achieve the best state representation at the equilibrium. This can be achieved by minimizing the relative entropy between our modeled state distribution and the true posterior distribution.

In **Precog**, we make better predictions by minimizing the cross-entropy between the actual and predictive output distributions.

In **Pathfinder**, we make the tradeoffs between the trained policy’s performance and robustness, similarly in the well-known exploitation and exploration tradeoffs in reinforcement learning. This is the same as balancing the cost entropy vs. the policy’s action-selection entropy. One goal of a desirable action policy is to prefer selecting actions for minimizing the mean and entropy of the long-term cost across different states and conditions. Thus, this policy can achieve high return and low risk outcomes in a dynamic environment. The other goal is to make the policy more robust in a noisy and shifting environment through randomized action selections for acquiring new data with unexplored inputs.

FlowOps technology delivers breakthrough results through using the data to build a predictive model for KPIs and associated entropy, and then using the model to optimize our action policy. We use the past data to best represent the current state of the dynamic process.

**Sentinel**extracts key patterns from the past sequence of signals and compresses them into the latent state.**Precog**predicts what is likely to happen and why it’s happening for any given condition or action.**Pathfinder**uses the Precog model for what-if simulations and learns the action policy to optimize the flow-at-risk and the entropy.

FlowOps solves the most representative problem for Enterprise AI in a causal setting: What would the **future sensor X(t+1)** and **target Y(t+1)** look like if we take an **action A(t)** in given **condition C(t)** for the current **latent state S(t)** of the process?

It is important that the **latent state** is a Bayesian probabilistic state so it’s being updated from the prior state distribution to the posterior state distribution after observing input-output data at each timestep. The latent state may be represented in a low-dimensional space so highly interpretable. It may capture shared and unique variances from correlated predictors to identify causal factors to generate **sensor** and **target** outputs. **Actions** are controllable inputs like planned supplies and productions, whereas **conditions** are uncontrollable inputs like SKU & DC locations, Covid-influenced mobility level, weather variables, holidays and so on. Both **sensors** and **targets** are measured outputs, and they are to be predicted over the future. In particular, **targets** are those variables to be optimized in **action** planning.

We first train our deep probabilistic model (refer to my previous blog, “Deep Probabilistic Decision Machines (DPDM) for building a causally generative process model-based action control in Enterprise AI”) using the historical data & knowledge, and then use it as a simulator for **action** planning.

We’re proud to be the leading provider in the Flow Operations category. It aligns with the Noodle.ai values. FlowOps is human-centered AI, built to minimize a human operator’s cognitive entropy and maximize the overall enterprise’s operational flow efficiency.

See Flow Operations and the breakthrough data science that fuel it, at live. Request a demo of our products for manufacturing and supply chain at Noodle.ai.

#### References:

[1] Doya, Kenji. “What are the computations of the cerebellum, the basal ganglia and the cerebral cortex?.” Neural networks 12.7-8 (1999): 961-974.

[2] Friston, Karl, et al. “Active inference and learning.” Neuroscience & Biobehavioral Reviews 68 (2016): 862-879