Reflexive Game Theory: Recursive Decision Modeling

• Reflexive Game Theory is a formal framework for modeling group decisions by embedding recursive self-modeling of agents’ strategies using Boolean algebra.
• The framework employs constructs like influence matrices, relationship graphs, and diagonal folding to capture hierarchical, multistage decision processes and dynamic learning.
• Its applications span human–robot systems, affective computing, and autonomous-agent collectives, offering insights into control scenarios and strategic reflexion beyond classical equilibria.

Reflexive Game Theory (RGT) is a formal framework for modeling and predicting the choices of autonomous subjects in group decision-making contexts, where each subject not only evaluates the objective alternatives but also hierarchically models how other subjects will choose, including how those subjects model each other’s intentions. The theory’s core innovation is the embedding of self-referential, hierarchical images—“I think that you think that I think …”—within the decision process, fully formalized via Boolean algebras and combinatorial logic. RGT generalizes both classical game-theoretic and collective-behavior models by explicitly structuring mutual reflexion, and has been systematically extended to handle multistage decision making, intention recognition dynamics, emotion modeling, learning, and control in human–robot and autonomous-agent collectives (Tarasenko, 2012, Novikov et al., 2018, Fujimoto et al., 2018, Tarasenko, 2010, Tarasenko, 2010, Tarasenko, 2015).

1. Mathematical Foundations of Reflexive Game Theory

RGT’s mathematical apparatus builds from several core constructs:

• Boolean Algebra of Alternatives: Let $X$ be the set of “elementary actions.” The full alternative set is $2^X$, endowed with union (+), intersection (·), and complement (¬). Each subject’s position is a subset $A \in 2^X$ (Tarasenko, 2012).
• Influence Matrix: For a group of $n$ subjects (indexed $i=1,\ldots,n$), the matrix $M$ with $(i,j)$ entry $M_{ij} \in 2^X$ encodes the influence of $i$ on $j$; $M_{ii} = x_i$ (the subject’s own point of view) (Tarasenko, 2012).
• Relationship Graph and Polynomial Representation: Subjects form vertices in a fully connected graph, with alliance ($\cdot$) and conflict (+) edges. This is converted to a polynomial $P$ in subject variables, e.g. $$P(x_1,\dots,x_n) = x_1 \cdot x_2 + x_3 \cdot (x_4 + x_5) + \ldots$$ (Tarasenko, 2012, Tarasenko, 2010).
• Diagonal Form and Folding: The polynomial $P$ is rewritten into a diagonal form that encodes nested reflexive layers. Diagonal-folding via $Q^W = Q + \overline{W}$ yields reflexive partitions, turning the graph into a hierarchy of decision equations (Tarasenko, 2010).
• Reflexive (Decision) Functions: Each agent $i$ is assigned a reflexive function $F_i(x_1, \ldots, x_n)$ derived from the folded polynomial, summarizing the combinatorics of influences shaping $i$’s choice (Tarasenko, 2012).

For each subject $i$, the canonical decision equation takes the form:

$$x_i = F_i(x_1, ..., x_n) \cdot x_i + G_i(x_1, ..., x_n) \cdot \neg x_i$$

The solution concept is the “decision interval” $F_i(x) \supseteq x_i \supseteq G_i(x)$, yielding all admissible choices given group structure and mutual influences (Tarasenko, 2012, Tarasenko, 2010).

2. Reflexion, Informational Structures, and Strategic Hierarchies

RGT formalizes reflexion through recursive belief and strategy hierarchies:

• Informational Reflexion: Agents possess awareness structures $I = \{\theta_\tau\}$, representing belief chains about states of nature and higher-order beliefs. A reflexive game is $\Gamma = (N, X_i, u_i(\cdot), I)$. Even when $I$ has finite complexity, agents' actions are dictated by their position in the awareness hierarchy (Novikov et al., 2018).
• Strategic Reflexion: Agents are partitioned into levels (e.g. level-$k$ in Level-$k$ or Cognitive Hierarchies models), each modeling lower ranks as less sophisticated. Each level assigns distinct response functions—best response or quantal response—against expected strategies, yielding a reflexive equilibrium $s^* = \{s_j^*\}$ (Novikov et al., 2018).
• Unified Reflexion Graphs: Combined informational and strategic reflexion is represented as directed graphs with real and phantom agents, where links encode both belief and strategic-rank relationships. Equilibrium is achieved as a profile of best responses at each graph node (Novikov et al., 2018).

Notably, under common knowledge, the RGT informational equilibrium collapses to the Nash equilibrium. However, RGT’s reflexive equilibrium can differ from Nash, as subjects best respond not only to realized choices but to their constructed images of all others’ decision processes (Novikov et al., 2018).

3. Multistage and Dynamic Reflexive Decision Processes

RGT generalizes to multistage and dynamic contexts (Tarasenko, 2012):

• Two-Stage Model: Decision making subdivides into a preparatory stage (parameters of the final decision process—group composition, structure, alternatives—are themselves set reflexively) and a final stage implementing the ultimate choice. Each is an RGT session (Tarasenko, 2012).
• Arbitrary Multistage Extension: The process iterates, updating polynomials and influence matrices at each stage via new RGT sessions, until all parameters stabilize. Parallel (batch) or consecutive (sequential) decisions are possible, allowing representation of complex multi-session organizational dynamics (Tarasenko, 2012).
• Dynamic Learning and Collective Behavior: RGT incorporates learning via belief-based (fictitious play), reinforcement-based (EWA), or hybrid learning models. Strategic reflexion propagates into time-dynamic schemes: $$x_i(t) = x_i(t-1) + \gamma_t(W_i(\text{predicted }x_{-i}(t)) - x_i(t-1))$$ (Novikov et al., 2018).
• Intention Recognition Dynamics: RGT has been formalized as coupled functional dynamics, where at each round $t$ each player $i$ maintains an intention function $f_i^{(t)}$, updated by recognition degrees $\alpha_i$, interpolating between classical Nash and Stackelberg equilibria and yielding new classes of $\alpha$-equilibria not accessible to traditional models (Fujimoto et al., 2018).

4. Computational Solution Methods and the Inverse Task

The RGT solution pipeline involves combinatorial manipulation in Boolean algebras and is algorithmically tractable for finite settings:

• Forward (Direct) Task: Given group structure and mutual influences, each subject solves its canonical decision equation, yielding its interval of admissible choices (Tarasenko, 2010).
• Inverse Task: The inverse task computes all influence assignments required to force a “controlled” subject to a target choice or set thereof, via a system of Boolean equations. If the target lies outside the admissible interval, the subject enters frustration, and control is impossible without modifying the group structure (Tarasenko, 2010).
• Algorithmic Schema: Influence equations are successively simplified via isolation of variables and reduction to interval solutions. For alliance and conflict operators, explicit interval formulas are available. The process is amenable to recursive implementation and hardware acceleration (Tarasenko, 2010).
• Extensions to Autonomous Agents: The solution architecture generalizes to groups containing both humans and robots, with additional modules (e.g., safety filters) encoding constraints such as Asimov's laws, and integrated interaction modules for mixed-agent control protocols (Tarasenko, 2010, Tarasenko, 2015).

5. Emotional and Affective Reflexive Games

RGT’s Boolean-algebraic machinery supports modeling emotional and affective reflexion:

• Emotional Reflexive Games (ERG): By encoding emotions as points in the PAD (Pleasure–Arousal–Dominance) model, with extremal states mapped to $3$-bit Boolean vectors, alliances and conflicts correspond to conjunctions and disjunctions of these vectors, yielding a formal model of affective group dynamics (Tarasenko, 2010).
• Hierarchies of Emotional Images: Diagonal-form folding propagates emotional states through hierarchies of reflexive images, capturing the fast, unconscious spread of affect through teams and collectives. ERG thus enables reflexive emotional steering protocols in human and hybrid groups (Tarasenko, 2010).

6. Applications in Autonomous Agents and Human–Robot Systems

RGT directly supports design and analysis of both human and autonomous-agent collectives:

• Communication Architecture: In robotic collectives, RGT-relevant information (alliances, conflicts, alternatives) is broadcast via frequency-division multiplexing, implemented with resonate-and-fire neural models to realize robust distributed group representation and influence assignment (Tarasenko, 2015).
• Control Scenarios: Case studies include robotic babysitters enforcing safe options, rescue robots orchestrating help via frustration induction, and cooperative strategies in resource allocation—all derived explicitly from RGT inference and inverse solution procedures (Tarasenko, 2010, Tarasenko, 2015).
• Adaptability: RGT-based systems flexibly switch between cooperative, competitive, and neutral configurations by dynamically reconfiguring relationship graphs and influence matrices via distributed cyclic communication protocols (Tarasenko, 2015).

7. Theoretical Significance, Limitations, and Extensions

The primary contributions of RGT are its explicit recursivity, capacity to capture both informational and strategic reflexion, and rigorous linkage to both classical and behavioral game theory (Novikov et al., 2018, Tarasenko, 2012). RGT is highly general: it nests Nash and Stackelberg equilibria as limiting cases but enables a much broader class of equilibrium behavior under partial recognition and hierarchical group structures (Fujimoto et al., 2018).

Limitations include the combinatorial growth of the solution space with number of agents and alternatives, and (as suggested by the experience with “super-active groups”) the existence of certain configurations where control is impossible without changing group topology (Tarasenko, 2010). The formalism is best suited to finite, well-structurable decision contexts; generalization to infinite, continuous, or highly stochastic settings remains an area for exploration.

Extensions to learning, evolutionary dynamics, and algorithmic behavioral models have been formalized for both static and dynamic games, providing a comprehensive toolkit for incorporating reflexion into a range of static and evolving multiagent decision problems, including domains in human–robot interaction, affective computing, distributed control, and political negotiation (Novikov et al., 2018, Tarasenko, 2010, Tarasenko, 2012, Tarasenko, 2015).