Algorithm 1 Perishable-Aware Route Optimization via Q-Learning with Context-Aware Weights and Conflict Avoidance
1:Input: Cold-chain graph $G(V,E)$, perishability profile P, disruption model D, emission matrix C   2:Initialize: Q-table $Q(s,a)\leftarrow 0$; learning rate $\alpha$; discount factor $\gamma$; exploration rate $ϵ$   3:Initialize: Static priority coefficients ${\alpha }_1,{\alpha }_2,{\alpha }_3$
4:Initialize: Shared intention buffer $\mathcal{B}\leftarrow \varnothing$	▹ For coordination
5:for each episode do   6:       Initialize joint global state $S_0=\{s_0^{\left(1\right)},\dots ,s_0^{\left(n\right)}\}$ using D   7:       while shipment not delivered do   8:             for each routing agent i do   9:                With probability $ϵ$, choose random action $a^{\left(i\right)}$ 10:              Otherwise, choose $a^{\left(i\right)}\leftarrow arg{max}_{a^{\prime }}Q(s^{\left(i\right)},a^{\prime })$
11:                Append $(s^{\left(i\right)},a^{\left(i\right)})$ to $\mathcal{B}$	▹ Declare action intention
12:             end for 13:             Detect conflicts in $\mathcal{B}$ (e.g., duplicate vehicle or route allocation) 14:             if conflict detected then 15:             Apply coordination penalty $\rho$ or reassign conflicting agent(s) via tie-breaking 16:             end if 17:             for each agent i do 18:             Execute $a^{\left(i\right)}$, observe $s^{\prime \left(i\right)}$, travel time $t^{\left(i\right)}$, temp deviation $\mathrm{\Delta }{T}^{\left(i\right)}$, emissions $e^{\left(i\right)}$ 19:             Compute spoilage risk: ${\sigma }^{\left(i\right)}\leftarrow f(P,\mathrm{\Delta }{T}^{\left(i\right)})$ 20:             Extract context vector: $ct{x}^{\left(i\right)}=[\mathrm{\Delta }{T}^{\left(i\right)},\mathrm{traffic},\mathrm{SLA}\mathrm{priority}]$ 21:              Compute dynamic weights: $${\omega }_j^{\left(i\right)}={\displaystyle \frac{{\alpha }_j·ct{x}_j^{\left(i\right)}}{{\sum }_{k=1}^3{\alpha }_k·ct{x}_k^{\left(i\right)}}},j=1,2,3$$ 22:              Compute context-aware reward: $$r^{\left(i\right)}=-({\omega }_1^{\left(i\right)}{t}^{\left(i\right)}+{\omega }_2^{\left(i\right)}{\sigma }^{\left(i\right)}+{\omega }_3^{\left(i\right)}{e}^{\left(i\right)})-\rho$$ 23:              Update Q-table: $$Q(s^{\left(i\right)},a^{\left(i\right)})\leftarrow Q(s^{\left(i\right)},a^{\left(i\right)})+\alpha \left[r^{\left(i\right)}+\gamma \underset{a^{\prime }}{max}Q(s^{\prime \left(i\right)},a^{\prime })-Q(s^{\left(i\right)},a^{\left(i\right)})\right]$$ 24:             Update state: $s^{\left(i\right)}\leftarrow s^{\prime \left(i\right)}$ 25:             end for 26:             Clear intention buffer: $\mathcal{B}\leftarrow \varnothing$ 27:    end while 28:end for 29:Output: Learned policies ${\pi }_i^{*}\left(s\right)=arg{max}_{a}Q(s,a)$ for all agents i