Fugit-Based Heuristics in Adaptive Learning

• Fugit-based heuristics are adaptive decision-making frameworks that employ fuzzy logic and IF–THEN rules to adjust strategies in time-varying environments.
• They integrate neuro-fuzzy network architectures and adaptive forgetting factors to translate continuous evidence and time into actionable decisions.
• Empirical studies show that these heuristics enhance convergence in multi-agent games and effectively model human stopping behavior in tasks like the Canoe-Motion Task.

Fugit-based heuristics designate a class of algorithmic decision-making and learning approaches where agents adapt their strategies or stopping criteria based on dynamic, often time-varying, rules that can be described by flexible, heuristic frameworks. These frameworks eschew traditional fixed boundaries or memoryless update rules, instead integrating human-interpretable constructs such as fuzzy logic categories, exponentially adaptively weighted observations, and IF–THEN rules to address uncertainty, nonstationarity, and adaptivity in sequential tasks and strategic interactions. Fugit-based heuristics underpin neuro-fuzzy models in decision-making and adaptive forgetting factor mechanisms in game-theoretic learning.

1. Fuzzy Logic and Neuro-Fuzzy Formulations for Time-Varying Decisions

Fugit-based heuristics leverage fuzzy-logic rules to model the temporal evolution of decision policies in continuous state spaces. In evidence accumulation tasks, states are parameterized by accumulated evidence $x$ and elapsed time $t$, with fuzzy membership functions (MFs) used to encode subjective verbal categories such as "small," "medium," or "large" $x$ and "short," "medium," or "long" $t$. Gaussian MFs are typically applied, e.g.,

$$\mu_{\text{small}}(x) = \exp\left(-\frac{(x - \mu_S)^2}{\sigma_S^2}\right)$$

Decisions are characterized by output fuzzy sets (e.g., CONTINUE or STOP), with the instantaneous “desire to stop” determined by a set of IF–THEN rules, such as:

• IF $x$ is Small AND $t$ is Short THEN Continue
• IF $x$ is Medium AND $t$ is Long THEN Stop

Rule-firing strengths, normalization, and center-average defuzzification mathematically formalize the moment-by-moment mapping $y(x, t)\colon (x, t) \rightarrow [0, 1]$, encoding the propensity to stop (Khodadadi et al., 2016). This approach captures the empirically observed time-varying stopping heuristics even in naïve subjects, suggesting innate or heuristically learned adaptation rather than explicit optimization.

2. Heuristic Weighting in Adaptive Game-Theoretic Play

In multi-agent potential games, fugit-based heuristics manifest as adaptive weightings of historical actions through schemes such as Adaptive Forgetting Factor Fictitious Play (AFFFP). Here, agents do not assume that opponents' strategies are stationary; rather, they apply a heuristic, exponentially weighted scheme to prioritize recent observations. This is distinct from classical fictitious play, which performs simple counts, or geometric fictitious play, which uses a fixed decay factor. The adaptive forgetting factor $\gamma_t \in (0,1]$ is updated online via stochastic gradient ascent on the streaming log-likelihood of observed actions:

$$w_{t+1}(a) = \gamma_t \cdot w_t(a) + I\{\text{opponent played } a \text{ at } t+1\}$$

where $w_t(a)$ denotes the current weight for opponent action $a$, with $\gamma_t$ being adjusted to maximize prediction accuracy (Smyrnakis et al., 2011). This enables real-time tracking of non-stationary environments via fugit-based adaptation.

3. Algorithmic Architectures and Learning Procedures

Fugit-based heuristics are realized in computational models with specialized architectures:

• Neuro-Fuzzy Model: The fuzzy decision system is mapped onto a four-layer feed-forward network—fuzzification computes MF activations, rule units compute conjunctions, aggregation merges rule outputs, and defuzzification delivers actionable scores. Learning uses supervised back-propagation to fit MF parameters to behavioral data and reinforcement learning (TD error) to update rule consequents, with eligibility traces for credit assignment:

$$c_i \leftarrow c_i + \alpha \cdot \delta_k \cdot z_i(k)$$

where $c_i$ are rule-consequent weights, $\delta_k$ is the temporal-difference error, and $z_i(k)$ is the eligibility trace per rule (Khodadadi et al., 2016).

• AFFFP: For each opponent, weights and auxiliary gradients per action are recursively updated; the forgetting factor is tuned using the gradient of the log-likelihood of the predicted versus observed action, projected to the interval $(0,1)$. Each agent forms mixed strategies based on current empirical priors and selects actions using best response or logit response mappings (Smyrnakis et al., 2011).

4. Experimental Paradigms and Empirical Performance

Fugit-based heuristics have been empirically validated in both human decision-making and decentralized optimization contexts:

• Canoe-Motion Task: Subjects must balance speed and accuracy to maximize reward. Behavioral data indicate that even untrained subjects use time-varying stopping heuristics, codified by fuzzy IF–THEN rules. Fast and slow responders differ in their rule-bases—fast subjects assign "stop" to smaller values of $(x, t)$, slow subjects require higher thresholds. With practice, MF centers shift, but rule-structure bias persists, suggesting that initial fugit-based heuristics steer subsequent learning dynamics. The neuro-fuzzy model achieves fit quality comparable to drift-diffusion models, with interpretable psychological parameters (Khodadadi et al., 2016).
• Potential Games: AFFFP has been benchmarked in three-player climbing hill games, vehicle-target assignment, and disaster-management ambulance allocation. AFFFP achieves faster or at least as reliable convergence to Nash equilibria as classical or geometric FP. For example, in the climbing hill game, AFFFP converged in approximately 100 iterations versus 300 for geometric FP, and in ambulance dispatch, achieved up to 20% more completions than geometric FP for $N_{\text{amb}}=20$, while matching or exceeding prior approaches and maintaining full decentralization (Smyrnakis et al., 2011).

5. Theoretical Properties and Interpretability

Fugit-based heuristics offer a trade-off: they are robust to nonstationarity and provide interpretable structure (verbal categories, IF–THEN rules, adaptive weightings) but often lack guarantees of optimality or formal convergence proofs beyond those inherited from their static or degenerate limits. For neuro-fuzzy models, parameter adjustment mimics learning under reward rate maximization, but initial rule-structure biases can persist. For AFFFP, convergence is empirically demonstrated but not formally guaranteed; whenever the forgetting factor stabilizes, classical convergence properties are recovered (Khodadadi et al., 2016, Smyrnakis et al., 2011).

A distinguishing feature is psychological and practical interpretability. Fuzzy MFs map directly to human-relevant concepts ("small," "medium," "long"), and rule bases encode qualitative strategies, in contrast to less transparent parameterizations in classical models.

6. Relation to Broader Research and Applications

Fugit-based heuristics connect linguistic rule-based decision paradigms in neuroscience and psychology with adaptive, decentralized learning in multi-agent optimization. They are applicable to domains requiring ongoing adaptation, including sensor networks, disaster relief, traffic scheduling, as well as laboratory tasks investigating the speed–accuracy trade-off in human subjects. Their flexibility and interpretability make them well-suited for modeling both human strategic behavior and distributed algorithmic policies in uncertain and evolving environments (Khodadadi et al., 2016, Smyrnakis et al., 2011).