IndigoVX: Where Human Intelligence Meets AI for Optimal Decision Making

Feedback loops are a universal feature of artificial intelligence, underpinning well-established techniques such as gradient descent, genetic algorithms, and reinforcement learning, among others (Goodfellow et al, 2016;Russell and Norvig, 2016). These iterative processes leverage a continuous cycle of adjustment and refinement to guide a system towards a desired outcome (Domingos, 2015). We introduce a novel approach that augments a human expert with an AI to achieve an optimal plan. Our approach is widely applicable, for example to collaborative business strategies, game playing, creative writing, or even planning a route for a road trip. We consider the planning problem as a classic multi-objective optimization problem, given weights and scores (Deb, 2001). The scores reflect the fitness of a candidate plan for a goal, and the weights reflect the relative importance of each scoring function: In the field of automated or machine learning-assisted decision-making, the essential aspect of uncertainty is often overlooked (Kochenderfer, 2015). In Indigo, working together as a collaborative team, the human expert provides context-specific expertise and judgment, while the AI provides data-driven insights, scalability, and iterative problem-solving capacity. Human supervision can reduce uncertainty by providing additional context (Holzinger, 2016).

The approach draws on and situates itself within several strands of prior work. It is inspired by Reinforcement Learning (RL), where iterative improvement via policies and rewards guides behavior; here, the human expert serves as a proxy for the environment, enabling policy-like iteration without a fully specified MDP (Sutton and Barto, 2018; Christiano et al., 2017). The method acknowledges uncertainty in decision-making and highlights the role of human-in-the-loop systems to provide context and reduce uncertainty (Kochenderfer, 2015; Holzinger, 2016). In multi-objective optimization, the use of weighted sums and Pareto concepts motivates IndigoVX’s dynamic weighting of three normalized scores, aligning with classic and interactive multiobjective optimization principles while noting potential biases of static weights (Deb, 2001; Marler and Arora, 2004; Miettinen and Mäkelä, 2006). The paper also relates IndigoVX to traditional decision-making matrices—such as Ease/Effect, Iron Triangle, and Eisenhower Matrix—by showing how their qualitative dimensions can be mapped into the three-score quantitative schema (Boardman et al., 2018; Atkinson, 1999; Covey, 2020). Further, it connects to work on collective intelligence and ensemble decision-making, suggesting potential gains from combining human and multiple AI participants with voting (Malone, 2018; Surowiecki, 2004; Woolley et al., 2010).

The IndigoVX algorithm is a human-AI collaborative optimization loop designed for well-defined, specific, and quantifiable goals. Core components: - Objective: A clearly specified, measurable goal. - Participants: One AI system and one human expert (or an ensemble configuration). Initialization: - The AI and human expert jointly draft an initial plan as the baseline for subsequent iterations. Scoring schema: - Define three evaluation scores for the plan, each quantized on a 0.5 increment scale from 0 to 10. These scores assess plan quality and effectiveness with respect to the goal. Optimization loop (iterative steps): 1) Rate the current plan using the agreed three-score schema. 2) Propose a list of concrete edits to the plan, explaining how each edit is expected to improve the scores. Convergence criterion: - Continue iterations until the weighted score differences across all three scores from one iteration to the next fall below a threshold (e.g., 0.5) over a window of several prior iterations. The process is intended to follow an intuitive trajectory akin to steepest descent. Scoring scale rationale: - A 0–10 scale is familiar, cognitively accessible (with 0.5 increments providing nuanced choices like 7.5 vs 7 or 8), and includes a neutral midpoint at 5. The 0.5 step increases precision without overburdening raters. Dynamic weighting: - Three normalized scores are combined via a weighted sum. Weights are dynamic and can be adjusted throughout the loop to reflect shifting priorities or new insights from the human expert. This adaptive approach aims to mitigate biases of static weighted sums and better balance trade-offs across objectives, with future refinement via Pareto-based methods suggested. Ensembles and collective intelligence (optional extension): - Alternative architectures include multiple human experts with one AI, or one human with multiple AIs. A voting step can be added wherein each participant independently proposes edits and then all participants vote on the next move through parameter space. The paper notes practical success with a single human expert paired with capable LLMs (e.g., ChatGPT-4, Bard, Claude 2).

- Proposed IndigoVX, a simple, general human-AI collaborative loop for plan optimization around a quantified, three-score schema. - Defined a practical scoring framework: three scores on a 0–10 scale with 0.5 increments to balance familiarity, cognitive ease, and nuance, including a neutral midpoint at 5. - Established an iterative optimization process with explicit steps (scoring, proposing edits) and a convergence criterion based on a weighted score difference threshold (e.g., 0.5) across a window of iterations. - Introduced dynamic weighting of normalized scores to reflect changing priorities and mitigate biases of static weighted sums in multi-objective optimization; suggested future Pareto-based refinements. - Articulated benefits inspired by RL without requiring a fully specified environment: hybrid human-AI intelligence, integration of rich context, and applicability even when precise mathematical models are unavailable. - Outlined an ensemble extension using multiple AIs or experts with a voting mechanism to exploit collective intelligence while noting practical emphasis on pairing a single expert with capable LLMs.

IndigoVX addresses the planning and decision-making problem by combining human contextual judgment with AI-driven iteration and data-informed suggestions. The iterative scoring-and-editing loop functions analogously to policy iteration in RL, yet operates in broader, less formally defined environments by leveraging the human expert as an environment proxy. This design tackles uncertainty by explicitly incorporating human supervision and context, and it supports multi-objective trade-offs through a three-score schema with adjustable weights. The method’s quantified structure facilitates real-time adaptation: participants can reweight objectives as situational priorities evolve, enabling balanced progress toward an optimal or near-optimal plan. Mapping traditional qualitative decision matrices into the three-score framework connects IndigoVX to established practice while enhancing scalability and precision through AI support. The ensemble extension further broadens applicability by harnessing collective intelligence, though practical deployments may favor minimal yet high-capability teams. Overall, the framework provides a flexible, implementable pathway to improved decisions in domains like business strategy, games, and creative planning without requiring detailed environmental models.

This decade has arguably already seen the emergence of early-stage Artificial General Intelligence (proto-AGI). The authors envision futures that may include invasive neural augmentation via BCIs; against that backdrop, Indigo is presented as a simple, foundational approach to human-AI collaborative problem-solving and a non-invasive step toward hybrid intelligence. The algorithm is straightforward to implement—requiring little more than pen and paper plus access to strong natural language models—yet provides a structured, iterative pathway to optimize plans toward defined goals. Future directions include refining dynamic weighting with Pareto-based techniques, exploring ensemble voting mechanisms among multiple AIs and experts, and applying the method across diverse decision-making contexts.

- Scoring schema sensitivity: Different problems may require different scoring schemes; while flexibility and adaptivity are allowed, schema selection can influence outcomes and may require standardization for similar use cases. - Subjectivity: Score choices and evaluations can be partly subjective; the paper argues this can be a strength for capturing human perspective, but it also introduces variability and potential bias. - Convergence definition: The chosen threshold-based convergence (e.g., 0.5 difference over a window) is heuristic; it may not guarantee global optimality and depends on weighting and window selection. - Weighted sum bias: Traditional weighted sums can bias solutions toward Pareto front corners; the proposed dynamic weighting mitigates but does not eliminate this, and full Pareto-based refinement is left for future work. - Environmental formalization: Unlike RL in a well-defined environment, IndigoVX operates in abstract contexts, which can complicate objective measurement and reproducibility. - Ensemble costs: While collective intelligence can improve solutions, involving multiple AIs/experts introduces additional computational and coordination overhead.